]>
Commit | Line | Data |
---|---|---|
75bb4625 JA |
1 | /* |
2 | * Block multiqueue core code | |
3 | * | |
4 | * Copyright (C) 2013-2014 Jens Axboe | |
5 | * Copyright (C) 2013-2014 Christoph Hellwig | |
6 | */ | |
320ae51f JA |
7 | #include <linux/kernel.h> |
8 | #include <linux/module.h> | |
9 | #include <linux/backing-dev.h> | |
10 | #include <linux/bio.h> | |
11 | #include <linux/blkdev.h> | |
f75782e4 | 12 | #include <linux/kmemleak.h> |
320ae51f JA |
13 | #include <linux/mm.h> |
14 | #include <linux/init.h> | |
15 | #include <linux/slab.h> | |
16 | #include <linux/workqueue.h> | |
17 | #include <linux/smp.h> | |
18 | #include <linux/llist.h> | |
19 | #include <linux/list_sort.h> | |
20 | #include <linux/cpu.h> | |
21 | #include <linux/cache.h> | |
22 | #include <linux/sched/sysctl.h> | |
23 | #include <linux/delay.h> | |
aedcd72f | 24 | #include <linux/crash_dump.h> |
88c7b2b7 | 25 | #include <linux/prefetch.h> |
320ae51f JA |
26 | |
27 | #include <trace/events/block.h> | |
28 | ||
29 | #include <linux/blk-mq.h> | |
30 | #include "blk.h" | |
31 | #include "blk-mq.h" | |
32 | #include "blk-mq-tag.h" | |
33 | ||
34 | static DEFINE_MUTEX(all_q_mutex); | |
35 | static LIST_HEAD(all_q_list); | |
36 | ||
37 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx); | |
38 | ||
320ae51f JA |
39 | /* |
40 | * Check if any of the ctx's have pending work in this hardware queue | |
41 | */ | |
42 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) | |
43 | { | |
44 | unsigned int i; | |
45 | ||
569fd0ce | 46 | for (i = 0; i < hctx->ctx_map.size; i++) |
1429d7c9 | 47 | if (hctx->ctx_map.map[i].word) |
320ae51f JA |
48 | return true; |
49 | ||
50 | return false; | |
51 | } | |
52 | ||
1429d7c9 JA |
53 | static inline struct blk_align_bitmap *get_bm(struct blk_mq_hw_ctx *hctx, |
54 | struct blk_mq_ctx *ctx) | |
55 | { | |
56 | return &hctx->ctx_map.map[ctx->index_hw / hctx->ctx_map.bits_per_word]; | |
57 | } | |
58 | ||
59 | #define CTX_TO_BIT(hctx, ctx) \ | |
60 | ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1)) | |
61 | ||
320ae51f JA |
62 | /* |
63 | * Mark this ctx as having pending work in this hardware queue | |
64 | */ | |
65 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
66 | struct blk_mq_ctx *ctx) | |
67 | { | |
1429d7c9 JA |
68 | struct blk_align_bitmap *bm = get_bm(hctx, ctx); |
69 | ||
70 | if (!test_bit(CTX_TO_BIT(hctx, ctx), &bm->word)) | |
71 | set_bit(CTX_TO_BIT(hctx, ctx), &bm->word); | |
72 | } | |
73 | ||
74 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
75 | struct blk_mq_ctx *ctx) | |
76 | { | |
77 | struct blk_align_bitmap *bm = get_bm(hctx, ctx); | |
78 | ||
79 | clear_bit(CTX_TO_BIT(hctx, ctx), &bm->word); | |
320ae51f JA |
80 | } |
81 | ||
b4c6a028 | 82 | void blk_mq_freeze_queue_start(struct request_queue *q) |
43a5e4e2 | 83 | { |
4ecd4fef | 84 | int freeze_depth; |
cddd5d17 | 85 | |
4ecd4fef CH |
86 | freeze_depth = atomic_inc_return(&q->mq_freeze_depth); |
87 | if (freeze_depth == 1) { | |
3ef28e83 | 88 | percpu_ref_kill(&q->q_usage_counter); |
b94ec296 | 89 | blk_mq_run_hw_queues(q, false); |
cddd5d17 | 90 | } |
f3af020b | 91 | } |
b4c6a028 | 92 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start); |
f3af020b TH |
93 | |
94 | static void blk_mq_freeze_queue_wait(struct request_queue *q) | |
95 | { | |
3ef28e83 | 96 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
43a5e4e2 ML |
97 | } |
98 | ||
f3af020b TH |
99 | /* |
100 | * Guarantee no request is in use, so we can change any data structure of | |
101 | * the queue afterward. | |
102 | */ | |
3ef28e83 | 103 | void blk_freeze_queue(struct request_queue *q) |
f3af020b | 104 | { |
3ef28e83 DW |
105 | /* |
106 | * In the !blk_mq case we are only calling this to kill the | |
107 | * q_usage_counter, otherwise this increases the freeze depth | |
108 | * and waits for it to return to zero. For this reason there is | |
109 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
110 | * exported to drivers as the only user for unfreeze is blk_mq. | |
111 | */ | |
f3af020b TH |
112 | blk_mq_freeze_queue_start(q); |
113 | blk_mq_freeze_queue_wait(q); | |
114 | } | |
3ef28e83 DW |
115 | |
116 | void blk_mq_freeze_queue(struct request_queue *q) | |
117 | { | |
118 | /* | |
119 | * ...just an alias to keep freeze and unfreeze actions balanced | |
120 | * in the blk_mq_* namespace | |
121 | */ | |
122 | blk_freeze_queue(q); | |
123 | } | |
c761d96b | 124 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
f3af020b | 125 | |
b4c6a028 | 126 | void blk_mq_unfreeze_queue(struct request_queue *q) |
320ae51f | 127 | { |
4ecd4fef | 128 | int freeze_depth; |
320ae51f | 129 | |
4ecd4fef CH |
130 | freeze_depth = atomic_dec_return(&q->mq_freeze_depth); |
131 | WARN_ON_ONCE(freeze_depth < 0); | |
132 | if (!freeze_depth) { | |
3ef28e83 | 133 | percpu_ref_reinit(&q->q_usage_counter); |
320ae51f | 134 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 135 | } |
320ae51f | 136 | } |
b4c6a028 | 137 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 138 | |
aed3ea94 JA |
139 | void blk_mq_wake_waiters(struct request_queue *q) |
140 | { | |
141 | struct blk_mq_hw_ctx *hctx; | |
142 | unsigned int i; | |
143 | ||
144 | queue_for_each_hw_ctx(q, hctx, i) | |
145 | if (blk_mq_hw_queue_mapped(hctx)) | |
146 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
3fd5940c KB |
147 | |
148 | /* | |
149 | * If we are called because the queue has now been marked as | |
150 | * dying, we need to ensure that processes currently waiting on | |
151 | * the queue are notified as well. | |
152 | */ | |
153 | wake_up_all(&q->mq_freeze_wq); | |
aed3ea94 JA |
154 | } |
155 | ||
320ae51f JA |
156 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) |
157 | { | |
158 | return blk_mq_has_free_tags(hctx->tags); | |
159 | } | |
160 | EXPORT_SYMBOL(blk_mq_can_queue); | |
161 | ||
94eddfbe | 162 | static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx, |
cc6e3b10 MC |
163 | struct request *rq, int op, |
164 | unsigned int op_flags) | |
320ae51f | 165 | { |
94eddfbe | 166 | if (blk_queue_io_stat(q)) |
cc6e3b10 | 167 | op_flags |= REQ_IO_STAT; |
94eddfbe | 168 | |
af76e555 CH |
169 | INIT_LIST_HEAD(&rq->queuelist); |
170 | /* csd/requeue_work/fifo_time is initialized before use */ | |
171 | rq->q = q; | |
320ae51f | 172 | rq->mq_ctx = ctx; |
cc6e3b10 | 173 | req_set_op_attrs(rq, op, op_flags); |
af76e555 CH |
174 | /* do not touch atomic flags, it needs atomic ops against the timer */ |
175 | rq->cpu = -1; | |
af76e555 CH |
176 | INIT_HLIST_NODE(&rq->hash); |
177 | RB_CLEAR_NODE(&rq->rb_node); | |
af76e555 CH |
178 | rq->rq_disk = NULL; |
179 | rq->part = NULL; | |
3ee32372 | 180 | rq->start_time = jiffies; |
af76e555 CH |
181 | #ifdef CONFIG_BLK_CGROUP |
182 | rq->rl = NULL; | |
0fec08b4 | 183 | set_start_time_ns(rq); |
af76e555 CH |
184 | rq->io_start_time_ns = 0; |
185 | #endif | |
186 | rq->nr_phys_segments = 0; | |
187 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
188 | rq->nr_integrity_segments = 0; | |
189 | #endif | |
af76e555 CH |
190 | rq->special = NULL; |
191 | /* tag was already set */ | |
192 | rq->errors = 0; | |
af76e555 | 193 | |
6f4a1626 TB |
194 | rq->cmd = rq->__cmd; |
195 | ||
af76e555 CH |
196 | rq->extra_len = 0; |
197 | rq->sense_len = 0; | |
198 | rq->resid_len = 0; | |
199 | rq->sense = NULL; | |
200 | ||
af76e555 | 201 | INIT_LIST_HEAD(&rq->timeout_list); |
f6be4fb4 JA |
202 | rq->timeout = 0; |
203 | ||
af76e555 CH |
204 | rq->end_io = NULL; |
205 | rq->end_io_data = NULL; | |
206 | rq->next_rq = NULL; | |
207 | ||
d9d8c5c4 | 208 | ctx->rq_dispatched[rw_is_sync(op, op_flags)]++; |
320ae51f JA |
209 | } |
210 | ||
5dee8577 | 211 | static struct request * |
cc6e3b10 | 212 | __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int op, int op_flags) |
5dee8577 CH |
213 | { |
214 | struct request *rq; | |
215 | unsigned int tag; | |
216 | ||
cb96a42c | 217 | tag = blk_mq_get_tag(data); |
5dee8577 | 218 | if (tag != BLK_MQ_TAG_FAIL) { |
cb96a42c | 219 | rq = data->hctx->tags->rqs[tag]; |
5dee8577 | 220 | |
cb96a42c | 221 | if (blk_mq_tag_busy(data->hctx)) { |
5dee8577 | 222 | rq->cmd_flags = REQ_MQ_INFLIGHT; |
cb96a42c | 223 | atomic_inc(&data->hctx->nr_active); |
5dee8577 CH |
224 | } |
225 | ||
226 | rq->tag = tag; | |
cc6e3b10 | 227 | blk_mq_rq_ctx_init(data->q, data->ctx, rq, op, op_flags); |
5dee8577 CH |
228 | return rq; |
229 | } | |
230 | ||
231 | return NULL; | |
232 | } | |
233 | ||
6f3b0e8b CH |
234 | struct request *blk_mq_alloc_request(struct request_queue *q, int rw, |
235 | unsigned int flags) | |
320ae51f | 236 | { |
d852564f CH |
237 | struct blk_mq_ctx *ctx; |
238 | struct blk_mq_hw_ctx *hctx; | |
320ae51f | 239 | struct request *rq; |
cb96a42c | 240 | struct blk_mq_alloc_data alloc_data; |
a492f075 | 241 | int ret; |
320ae51f | 242 | |
6f3b0e8b | 243 | ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT); |
a492f075 JL |
244 | if (ret) |
245 | return ERR_PTR(ret); | |
320ae51f | 246 | |
d852564f CH |
247 | ctx = blk_mq_get_ctx(q); |
248 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
6f3b0e8b | 249 | blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx); |
d852564f | 250 | |
cc6e3b10 | 251 | rq = __blk_mq_alloc_request(&alloc_data, rw, 0); |
6f3b0e8b | 252 | if (!rq && !(flags & BLK_MQ_REQ_NOWAIT)) { |
d852564f CH |
253 | __blk_mq_run_hw_queue(hctx); |
254 | blk_mq_put_ctx(ctx); | |
255 | ||
256 | ctx = blk_mq_get_ctx(q); | |
257 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
6f3b0e8b | 258 | blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx); |
cc6e3b10 | 259 | rq = __blk_mq_alloc_request(&alloc_data, rw, 0); |
cb96a42c | 260 | ctx = alloc_data.ctx; |
d852564f CH |
261 | } |
262 | blk_mq_put_ctx(ctx); | |
c76541a9 | 263 | if (!rq) { |
3ef28e83 | 264 | blk_queue_exit(q); |
a492f075 | 265 | return ERR_PTR(-EWOULDBLOCK); |
c76541a9 | 266 | } |
0c4de0f3 CH |
267 | |
268 | rq->__data_len = 0; | |
269 | rq->__sector = (sector_t) -1; | |
270 | rq->bio = rq->biotail = NULL; | |
320ae51f JA |
271 | return rq; |
272 | } | |
4bb659b1 | 273 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 274 | |
1f5bd336 ML |
275 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, int rw, |
276 | unsigned int flags, unsigned int hctx_idx) | |
277 | { | |
278 | struct blk_mq_hw_ctx *hctx; | |
279 | struct blk_mq_ctx *ctx; | |
280 | struct request *rq; | |
281 | struct blk_mq_alloc_data alloc_data; | |
282 | int ret; | |
283 | ||
284 | /* | |
285 | * If the tag allocator sleeps we could get an allocation for a | |
286 | * different hardware context. No need to complicate the low level | |
287 | * allocator for this for the rare use case of a command tied to | |
288 | * a specific queue. | |
289 | */ | |
290 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT))) | |
291 | return ERR_PTR(-EINVAL); | |
292 | ||
293 | if (hctx_idx >= q->nr_hw_queues) | |
294 | return ERR_PTR(-EIO); | |
295 | ||
296 | ret = blk_queue_enter(q, true); | |
297 | if (ret) | |
298 | return ERR_PTR(ret); | |
299 | ||
300 | hctx = q->queue_hw_ctx[hctx_idx]; | |
301 | ctx = __blk_mq_get_ctx(q, cpumask_first(hctx->cpumask)); | |
302 | ||
303 | blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx); | |
304 | rq = __blk_mq_alloc_request(&alloc_data, rw, 0); | |
305 | if (!rq) { | |
306 | blk_queue_exit(q); | |
307 | return ERR_PTR(-EWOULDBLOCK); | |
308 | } | |
309 | ||
310 | return rq; | |
311 | } | |
312 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
313 | ||
320ae51f JA |
314 | static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx, |
315 | struct blk_mq_ctx *ctx, struct request *rq) | |
316 | { | |
317 | const int tag = rq->tag; | |
318 | struct request_queue *q = rq->q; | |
319 | ||
0d2602ca JA |
320 | if (rq->cmd_flags & REQ_MQ_INFLIGHT) |
321 | atomic_dec(&hctx->nr_active); | |
683d0e12 | 322 | rq->cmd_flags = 0; |
0d2602ca | 323 | |
af76e555 | 324 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); |
0d2602ca | 325 | blk_mq_put_tag(hctx, tag, &ctx->last_tag); |
3ef28e83 | 326 | blk_queue_exit(q); |
320ae51f JA |
327 | } |
328 | ||
7c7f2f2b | 329 | void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq) |
320ae51f JA |
330 | { |
331 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
320ae51f JA |
332 | |
333 | ctx->rq_completed[rq_is_sync(rq)]++; | |
320ae51f | 334 | __blk_mq_free_request(hctx, ctx, rq); |
7c7f2f2b JA |
335 | |
336 | } | |
337 | EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request); | |
338 | ||
339 | void blk_mq_free_request(struct request *rq) | |
340 | { | |
341 | struct blk_mq_hw_ctx *hctx; | |
342 | struct request_queue *q = rq->q; | |
343 | ||
344 | hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu); | |
345 | blk_mq_free_hctx_request(hctx, rq); | |
320ae51f | 346 | } |
1a3b595a | 347 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 348 | |
c8a446ad | 349 | inline void __blk_mq_end_request(struct request *rq, int error) |
320ae51f | 350 | { |
0d11e6ac ML |
351 | blk_account_io_done(rq); |
352 | ||
91b63639 | 353 | if (rq->end_io) { |
320ae51f | 354 | rq->end_io(rq, error); |
91b63639 CH |
355 | } else { |
356 | if (unlikely(blk_bidi_rq(rq))) | |
357 | blk_mq_free_request(rq->next_rq); | |
320ae51f | 358 | blk_mq_free_request(rq); |
91b63639 | 359 | } |
320ae51f | 360 | } |
c8a446ad | 361 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 362 | |
c8a446ad | 363 | void blk_mq_end_request(struct request *rq, int error) |
63151a44 CH |
364 | { |
365 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
366 | BUG(); | |
c8a446ad | 367 | __blk_mq_end_request(rq, error); |
63151a44 | 368 | } |
c8a446ad | 369 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 370 | |
30a91cb4 | 371 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 372 | { |
3d6efbf6 | 373 | struct request *rq = data; |
320ae51f | 374 | |
30a91cb4 | 375 | rq->q->softirq_done_fn(rq); |
320ae51f | 376 | } |
320ae51f | 377 | |
ed851860 | 378 | static void blk_mq_ipi_complete_request(struct request *rq) |
320ae51f JA |
379 | { |
380 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
38535201 | 381 | bool shared = false; |
320ae51f JA |
382 | int cpu; |
383 | ||
38535201 | 384 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { |
30a91cb4 CH |
385 | rq->q->softirq_done_fn(rq); |
386 | return; | |
387 | } | |
320ae51f JA |
388 | |
389 | cpu = get_cpu(); | |
38535201 CH |
390 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) |
391 | shared = cpus_share_cache(cpu, ctx->cpu); | |
392 | ||
393 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
30a91cb4 | 394 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
395 | rq->csd.info = rq; |
396 | rq->csd.flags = 0; | |
c46fff2a | 397 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 398 | } else { |
30a91cb4 | 399 | rq->q->softirq_done_fn(rq); |
3d6efbf6 | 400 | } |
320ae51f JA |
401 | put_cpu(); |
402 | } | |
30a91cb4 | 403 | |
1fa8cc52 | 404 | static void __blk_mq_complete_request(struct request *rq) |
ed851860 JA |
405 | { |
406 | struct request_queue *q = rq->q; | |
407 | ||
408 | if (!q->softirq_done_fn) | |
c8a446ad | 409 | blk_mq_end_request(rq, rq->errors); |
ed851860 JA |
410 | else |
411 | blk_mq_ipi_complete_request(rq); | |
412 | } | |
413 | ||
30a91cb4 CH |
414 | /** |
415 | * blk_mq_complete_request - end I/O on a request | |
416 | * @rq: the request being processed | |
417 | * | |
418 | * Description: | |
419 | * Ends all I/O on a request. It does not handle partial completions. | |
420 | * The actual completion happens out-of-order, through a IPI handler. | |
421 | **/ | |
f4829a9b | 422 | void blk_mq_complete_request(struct request *rq, int error) |
30a91cb4 | 423 | { |
95f09684 JA |
424 | struct request_queue *q = rq->q; |
425 | ||
426 | if (unlikely(blk_should_fake_timeout(q))) | |
30a91cb4 | 427 | return; |
f4829a9b CH |
428 | if (!blk_mark_rq_complete(rq)) { |
429 | rq->errors = error; | |
ed851860 | 430 | __blk_mq_complete_request(rq); |
f4829a9b | 431 | } |
30a91cb4 CH |
432 | } |
433 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 434 | |
973c0191 KB |
435 | int blk_mq_request_started(struct request *rq) |
436 | { | |
437 | return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
438 | } | |
439 | EXPORT_SYMBOL_GPL(blk_mq_request_started); | |
440 | ||
e2490073 | 441 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
442 | { |
443 | struct request_queue *q = rq->q; | |
444 | ||
445 | trace_block_rq_issue(q, rq); | |
446 | ||
742ee69b | 447 | rq->resid_len = blk_rq_bytes(rq); |
91b63639 CH |
448 | if (unlikely(blk_bidi_rq(rq))) |
449 | rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq); | |
742ee69b | 450 | |
2b8393b4 | 451 | blk_add_timer(rq); |
87ee7b11 | 452 | |
538b7534 JA |
453 | /* |
454 | * Ensure that ->deadline is visible before set the started | |
455 | * flag and clear the completed flag. | |
456 | */ | |
457 | smp_mb__before_atomic(); | |
458 | ||
87ee7b11 JA |
459 | /* |
460 | * Mark us as started and clear complete. Complete might have been | |
461 | * set if requeue raced with timeout, which then marked it as | |
462 | * complete. So be sure to clear complete again when we start | |
463 | * the request, otherwise we'll ignore the completion event. | |
464 | */ | |
4b570521 JA |
465 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
466 | set_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
467 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
468 | clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags); | |
49f5baa5 CH |
469 | |
470 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
471 | /* | |
472 | * Make sure space for the drain appears. We know we can do | |
473 | * this because max_hw_segments has been adjusted to be one | |
474 | * fewer than the device can handle. | |
475 | */ | |
476 | rq->nr_phys_segments++; | |
477 | } | |
320ae51f | 478 | } |
e2490073 | 479 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 480 | |
ed0791b2 | 481 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
482 | { |
483 | struct request_queue *q = rq->q; | |
484 | ||
485 | trace_block_rq_requeue(q, rq); | |
49f5baa5 | 486 | |
e2490073 CH |
487 | if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
488 | if (q->dma_drain_size && blk_rq_bytes(rq)) | |
489 | rq->nr_phys_segments--; | |
490 | } | |
320ae51f JA |
491 | } |
492 | ||
ed0791b2 CH |
493 | void blk_mq_requeue_request(struct request *rq) |
494 | { | |
ed0791b2 | 495 | __blk_mq_requeue_request(rq); |
ed0791b2 | 496 | |
ed0791b2 | 497 | BUG_ON(blk_queued_rq(rq)); |
6fca6a61 | 498 | blk_mq_add_to_requeue_list(rq, true); |
ed0791b2 CH |
499 | } |
500 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
501 | ||
6fca6a61 CH |
502 | static void blk_mq_requeue_work(struct work_struct *work) |
503 | { | |
504 | struct request_queue *q = | |
2849450a | 505 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
506 | LIST_HEAD(rq_list); |
507 | struct request *rq, *next; | |
508 | unsigned long flags; | |
509 | ||
510 | spin_lock_irqsave(&q->requeue_lock, flags); | |
511 | list_splice_init(&q->requeue_list, &rq_list); | |
512 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
513 | ||
514 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
515 | if (!(rq->cmd_flags & REQ_SOFTBARRIER)) | |
516 | continue; | |
517 | ||
518 | rq->cmd_flags &= ~REQ_SOFTBARRIER; | |
519 | list_del_init(&rq->queuelist); | |
520 | blk_mq_insert_request(rq, true, false, false); | |
521 | } | |
522 | ||
523 | while (!list_empty(&rq_list)) { | |
524 | rq = list_entry(rq_list.next, struct request, queuelist); | |
525 | list_del_init(&rq->queuelist); | |
526 | blk_mq_insert_request(rq, false, false, false); | |
527 | } | |
528 | ||
8b957415 JA |
529 | /* |
530 | * Use the start variant of queue running here, so that running | |
531 | * the requeue work will kick stopped queues. | |
532 | */ | |
533 | blk_mq_start_hw_queues(q); | |
6fca6a61 CH |
534 | } |
535 | ||
536 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head) | |
537 | { | |
538 | struct request_queue *q = rq->q; | |
539 | unsigned long flags; | |
540 | ||
541 | /* | |
542 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
543 | * request head insertation from the workqueue. | |
544 | */ | |
545 | BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER); | |
546 | ||
547 | spin_lock_irqsave(&q->requeue_lock, flags); | |
548 | if (at_head) { | |
549 | rq->cmd_flags |= REQ_SOFTBARRIER; | |
550 | list_add(&rq->queuelist, &q->requeue_list); | |
551 | } else { | |
552 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
553 | } | |
554 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
555 | } | |
556 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
557 | ||
c68ed59f KB |
558 | void blk_mq_cancel_requeue_work(struct request_queue *q) |
559 | { | |
2849450a | 560 | cancel_delayed_work_sync(&q->requeue_work); |
c68ed59f KB |
561 | } |
562 | EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work); | |
563 | ||
6fca6a61 CH |
564 | void blk_mq_kick_requeue_list(struct request_queue *q) |
565 | { | |
2849450a | 566 | kblockd_schedule_delayed_work(&q->requeue_work, 0); |
6fca6a61 CH |
567 | } |
568 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
569 | ||
2849450a MS |
570 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
571 | unsigned long msecs) | |
572 | { | |
573 | kblockd_schedule_delayed_work(&q->requeue_work, | |
574 | msecs_to_jiffies(msecs)); | |
575 | } | |
576 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
577 | ||
1885b24d JA |
578 | void blk_mq_abort_requeue_list(struct request_queue *q) |
579 | { | |
580 | unsigned long flags; | |
581 | LIST_HEAD(rq_list); | |
582 | ||
583 | spin_lock_irqsave(&q->requeue_lock, flags); | |
584 | list_splice_init(&q->requeue_list, &rq_list); | |
585 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
586 | ||
587 | while (!list_empty(&rq_list)) { | |
588 | struct request *rq; | |
589 | ||
590 | rq = list_first_entry(&rq_list, struct request, queuelist); | |
591 | list_del_init(&rq->queuelist); | |
592 | rq->errors = -EIO; | |
593 | blk_mq_end_request(rq, rq->errors); | |
594 | } | |
595 | } | |
596 | EXPORT_SYMBOL(blk_mq_abort_requeue_list); | |
597 | ||
0e62f51f JA |
598 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
599 | { | |
88c7b2b7 JA |
600 | if (tag < tags->nr_tags) { |
601 | prefetch(tags->rqs[tag]); | |
4ee86bab | 602 | return tags->rqs[tag]; |
88c7b2b7 | 603 | } |
4ee86bab HR |
604 | |
605 | return NULL; | |
24d2f903 CH |
606 | } |
607 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
608 | ||
320ae51f | 609 | struct blk_mq_timeout_data { |
46f92d42 CH |
610 | unsigned long next; |
611 | unsigned int next_set; | |
320ae51f JA |
612 | }; |
613 | ||
90415837 | 614 | void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 615 | { |
46f92d42 CH |
616 | struct blk_mq_ops *ops = req->q->mq_ops; |
617 | enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER; | |
87ee7b11 JA |
618 | |
619 | /* | |
620 | * We know that complete is set at this point. If STARTED isn't set | |
621 | * anymore, then the request isn't active and the "timeout" should | |
622 | * just be ignored. This can happen due to the bitflag ordering. | |
623 | * Timeout first checks if STARTED is set, and if it is, assumes | |
624 | * the request is active. But if we race with completion, then | |
625 | * we both flags will get cleared. So check here again, and ignore | |
626 | * a timeout event with a request that isn't active. | |
627 | */ | |
46f92d42 CH |
628 | if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags)) |
629 | return; | |
87ee7b11 | 630 | |
46f92d42 | 631 | if (ops->timeout) |
0152fb6b | 632 | ret = ops->timeout(req, reserved); |
46f92d42 CH |
633 | |
634 | switch (ret) { | |
635 | case BLK_EH_HANDLED: | |
636 | __blk_mq_complete_request(req); | |
637 | break; | |
638 | case BLK_EH_RESET_TIMER: | |
639 | blk_add_timer(req); | |
640 | blk_clear_rq_complete(req); | |
641 | break; | |
642 | case BLK_EH_NOT_HANDLED: | |
643 | break; | |
644 | default: | |
645 | printk(KERN_ERR "block: bad eh return: %d\n", ret); | |
646 | break; | |
647 | } | |
87ee7b11 | 648 | } |
5b3f25fc | 649 | |
81481eb4 CH |
650 | static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
651 | struct request *rq, void *priv, bool reserved) | |
652 | { | |
653 | struct blk_mq_timeout_data *data = priv; | |
87ee7b11 | 654 | |
eb130dbf KB |
655 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
656 | /* | |
657 | * If a request wasn't started before the queue was | |
658 | * marked dying, kill it here or it'll go unnoticed. | |
659 | */ | |
a59e0f57 KB |
660 | if (unlikely(blk_queue_dying(rq->q))) { |
661 | rq->errors = -EIO; | |
662 | blk_mq_end_request(rq, rq->errors); | |
663 | } | |
46f92d42 | 664 | return; |
eb130dbf | 665 | } |
87ee7b11 | 666 | |
46f92d42 CH |
667 | if (time_after_eq(jiffies, rq->deadline)) { |
668 | if (!blk_mark_rq_complete(rq)) | |
0152fb6b | 669 | blk_mq_rq_timed_out(rq, reserved); |
46f92d42 CH |
670 | } else if (!data->next_set || time_after(data->next, rq->deadline)) { |
671 | data->next = rq->deadline; | |
672 | data->next_set = 1; | |
673 | } | |
87ee7b11 JA |
674 | } |
675 | ||
287922eb | 676 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 677 | { |
287922eb CH |
678 | struct request_queue *q = |
679 | container_of(work, struct request_queue, timeout_work); | |
81481eb4 CH |
680 | struct blk_mq_timeout_data data = { |
681 | .next = 0, | |
682 | .next_set = 0, | |
683 | }; | |
81481eb4 | 684 | int i; |
320ae51f | 685 | |
71f79fb3 GKB |
686 | /* A deadlock might occur if a request is stuck requiring a |
687 | * timeout at the same time a queue freeze is waiting | |
688 | * completion, since the timeout code would not be able to | |
689 | * acquire the queue reference here. | |
690 | * | |
691 | * That's why we don't use blk_queue_enter here; instead, we use | |
692 | * percpu_ref_tryget directly, because we need to be able to | |
693 | * obtain a reference even in the short window between the queue | |
694 | * starting to freeze, by dropping the first reference in | |
695 | * blk_mq_freeze_queue_start, and the moment the last request is | |
696 | * consumed, marked by the instant q_usage_counter reaches | |
697 | * zero. | |
698 | */ | |
699 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
700 | return; |
701 | ||
0bf6cd5b | 702 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data); |
320ae51f | 703 | |
81481eb4 CH |
704 | if (data.next_set) { |
705 | data.next = blk_rq_timeout(round_jiffies_up(data.next)); | |
706 | mod_timer(&q->timeout, data.next); | |
0d2602ca | 707 | } else { |
0bf6cd5b CH |
708 | struct blk_mq_hw_ctx *hctx; |
709 | ||
f054b56c ML |
710 | queue_for_each_hw_ctx(q, hctx, i) { |
711 | /* the hctx may be unmapped, so check it here */ | |
712 | if (blk_mq_hw_queue_mapped(hctx)) | |
713 | blk_mq_tag_idle(hctx); | |
714 | } | |
0d2602ca | 715 | } |
287922eb | 716 | blk_queue_exit(q); |
320ae51f JA |
717 | } |
718 | ||
719 | /* | |
720 | * Reverse check our software queue for entries that we could potentially | |
721 | * merge with. Currently includes a hand-wavy stop count of 8, to not spend | |
722 | * too much time checking for merges. | |
723 | */ | |
724 | static bool blk_mq_attempt_merge(struct request_queue *q, | |
725 | struct blk_mq_ctx *ctx, struct bio *bio) | |
726 | { | |
727 | struct request *rq; | |
728 | int checked = 8; | |
729 | ||
730 | list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) { | |
731 | int el_ret; | |
732 | ||
733 | if (!checked--) | |
734 | break; | |
735 | ||
736 | if (!blk_rq_merge_ok(rq, bio)) | |
737 | continue; | |
738 | ||
739 | el_ret = blk_try_merge(rq, bio); | |
740 | if (el_ret == ELEVATOR_BACK_MERGE) { | |
741 | if (bio_attempt_back_merge(q, rq, bio)) { | |
742 | ctx->rq_merged++; | |
743 | return true; | |
744 | } | |
745 | break; | |
746 | } else if (el_ret == ELEVATOR_FRONT_MERGE) { | |
747 | if (bio_attempt_front_merge(q, rq, bio)) { | |
748 | ctx->rq_merged++; | |
749 | return true; | |
750 | } | |
751 | break; | |
752 | } | |
753 | } | |
754 | ||
755 | return false; | |
756 | } | |
757 | ||
1429d7c9 JA |
758 | /* |
759 | * Process software queues that have been marked busy, splicing them | |
760 | * to the for-dispatch | |
761 | */ | |
762 | static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) | |
763 | { | |
764 | struct blk_mq_ctx *ctx; | |
765 | int i; | |
766 | ||
569fd0ce | 767 | for (i = 0; i < hctx->ctx_map.size; i++) { |
1429d7c9 JA |
768 | struct blk_align_bitmap *bm = &hctx->ctx_map.map[i]; |
769 | unsigned int off, bit; | |
770 | ||
771 | if (!bm->word) | |
772 | continue; | |
773 | ||
774 | bit = 0; | |
775 | off = i * hctx->ctx_map.bits_per_word; | |
776 | do { | |
777 | bit = find_next_bit(&bm->word, bm->depth, bit); | |
778 | if (bit >= bm->depth) | |
779 | break; | |
780 | ||
781 | ctx = hctx->ctxs[bit + off]; | |
782 | clear_bit(bit, &bm->word); | |
783 | spin_lock(&ctx->lock); | |
784 | list_splice_tail_init(&ctx->rq_list, list); | |
785 | spin_unlock(&ctx->lock); | |
786 | ||
787 | bit++; | |
788 | } while (1); | |
789 | } | |
790 | } | |
791 | ||
703fd1c0 JA |
792 | static inline unsigned int queued_to_index(unsigned int queued) |
793 | { | |
794 | if (!queued) | |
795 | return 0; | |
796 | ||
797 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); | |
798 | } | |
799 | ||
320ae51f JA |
800 | /* |
801 | * Run this hardware queue, pulling any software queues mapped to it in. | |
802 | * Note that this function currently has various problems around ordering | |
803 | * of IO. In particular, we'd like FIFO behaviour on handling existing | |
804 | * items on the hctx->dispatch list. Ignore that for now. | |
805 | */ | |
806 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) | |
807 | { | |
808 | struct request_queue *q = hctx->queue; | |
320ae51f JA |
809 | struct request *rq; |
810 | LIST_HEAD(rq_list); | |
74c45052 JA |
811 | LIST_HEAD(driver_list); |
812 | struct list_head *dptr; | |
1429d7c9 | 813 | int queued; |
320ae51f | 814 | |
5d12f905 | 815 | if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state))) |
320ae51f JA |
816 | return; |
817 | ||
0e87e58b JA |
818 | WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && |
819 | cpu_online(hctx->next_cpu)); | |
820 | ||
320ae51f JA |
821 | hctx->run++; |
822 | ||
823 | /* | |
824 | * Touch any software queue that has pending entries. | |
825 | */ | |
1429d7c9 | 826 | flush_busy_ctxs(hctx, &rq_list); |
320ae51f JA |
827 | |
828 | /* | |
829 | * If we have previous entries on our dispatch list, grab them | |
830 | * and stuff them at the front for more fair dispatch. | |
831 | */ | |
832 | if (!list_empty_careful(&hctx->dispatch)) { | |
833 | spin_lock(&hctx->lock); | |
834 | if (!list_empty(&hctx->dispatch)) | |
835 | list_splice_init(&hctx->dispatch, &rq_list); | |
836 | spin_unlock(&hctx->lock); | |
837 | } | |
838 | ||
74c45052 JA |
839 | /* |
840 | * Start off with dptr being NULL, so we start the first request | |
841 | * immediately, even if we have more pending. | |
842 | */ | |
843 | dptr = NULL; | |
844 | ||
320ae51f JA |
845 | /* |
846 | * Now process all the entries, sending them to the driver. | |
847 | */ | |
1429d7c9 | 848 | queued = 0; |
320ae51f | 849 | while (!list_empty(&rq_list)) { |
74c45052 | 850 | struct blk_mq_queue_data bd; |
320ae51f JA |
851 | int ret; |
852 | ||
853 | rq = list_first_entry(&rq_list, struct request, queuelist); | |
854 | list_del_init(&rq->queuelist); | |
320ae51f | 855 | |
74c45052 JA |
856 | bd.rq = rq; |
857 | bd.list = dptr; | |
858 | bd.last = list_empty(&rq_list); | |
859 | ||
860 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
320ae51f JA |
861 | switch (ret) { |
862 | case BLK_MQ_RQ_QUEUE_OK: | |
863 | queued++; | |
52b9c330 | 864 | break; |
320ae51f | 865 | case BLK_MQ_RQ_QUEUE_BUSY: |
320ae51f | 866 | list_add(&rq->queuelist, &rq_list); |
ed0791b2 | 867 | __blk_mq_requeue_request(rq); |
320ae51f JA |
868 | break; |
869 | default: | |
870 | pr_err("blk-mq: bad return on queue: %d\n", ret); | |
320ae51f | 871 | case BLK_MQ_RQ_QUEUE_ERROR: |
1e93b8c2 | 872 | rq->errors = -EIO; |
c8a446ad | 873 | blk_mq_end_request(rq, rq->errors); |
320ae51f JA |
874 | break; |
875 | } | |
876 | ||
877 | if (ret == BLK_MQ_RQ_QUEUE_BUSY) | |
878 | break; | |
74c45052 JA |
879 | |
880 | /* | |
881 | * We've done the first request. If we have more than 1 | |
882 | * left in the list, set dptr to defer issue. | |
883 | */ | |
884 | if (!dptr && rq_list.next != rq_list.prev) | |
885 | dptr = &driver_list; | |
320ae51f JA |
886 | } |
887 | ||
703fd1c0 | 888 | hctx->dispatched[queued_to_index(queued)]++; |
320ae51f JA |
889 | |
890 | /* | |
891 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
892 | * that is where we will continue on next queue run. | |
893 | */ | |
894 | if (!list_empty(&rq_list)) { | |
895 | spin_lock(&hctx->lock); | |
896 | list_splice(&rq_list, &hctx->dispatch); | |
897 | spin_unlock(&hctx->lock); | |
9ba52e58 SL |
898 | /* |
899 | * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but | |
900 | * it's possible the queue is stopped and restarted again | |
901 | * before this. Queue restart will dispatch requests. And since | |
902 | * requests in rq_list aren't added into hctx->dispatch yet, | |
903 | * the requests in rq_list might get lost. | |
904 | * | |
905 | * blk_mq_run_hw_queue() already checks the STOPPED bit | |
906 | **/ | |
907 | blk_mq_run_hw_queue(hctx, true); | |
320ae51f JA |
908 | } |
909 | } | |
910 | ||
506e931f JA |
911 | /* |
912 | * It'd be great if the workqueue API had a way to pass | |
913 | * in a mask and had some smarts for more clever placement. | |
914 | * For now we just round-robin here, switching for every | |
915 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
916 | */ | |
917 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
918 | { | |
b657d7e6 CH |
919 | if (hctx->queue->nr_hw_queues == 1) |
920 | return WORK_CPU_UNBOUND; | |
506e931f JA |
921 | |
922 | if (--hctx->next_cpu_batch <= 0) { | |
b657d7e6 | 923 | int cpu = hctx->next_cpu, next_cpu; |
506e931f JA |
924 | |
925 | next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask); | |
926 | if (next_cpu >= nr_cpu_ids) | |
927 | next_cpu = cpumask_first(hctx->cpumask); | |
928 | ||
929 | hctx->next_cpu = next_cpu; | |
930 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
b657d7e6 CH |
931 | |
932 | return cpu; | |
506e931f JA |
933 | } |
934 | ||
b657d7e6 | 935 | return hctx->next_cpu; |
506e931f JA |
936 | } |
937 | ||
320ae51f JA |
938 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
939 | { | |
19c66e59 ML |
940 | if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) || |
941 | !blk_mq_hw_queue_mapped(hctx))) | |
320ae51f JA |
942 | return; |
943 | ||
398205b8 | 944 | if (!async) { |
2a90d4aa PB |
945 | int cpu = get_cpu(); |
946 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 947 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 948 | put_cpu(); |
398205b8 PB |
949 | return; |
950 | } | |
e4043dcf | 951 | |
2a90d4aa | 952 | put_cpu(); |
e4043dcf | 953 | } |
398205b8 | 954 | |
27489a3c | 955 | kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work); |
320ae51f JA |
956 | } |
957 | ||
b94ec296 | 958 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
959 | { |
960 | struct blk_mq_hw_ctx *hctx; | |
961 | int i; | |
962 | ||
963 | queue_for_each_hw_ctx(q, hctx, i) { | |
964 | if ((!blk_mq_hctx_has_pending(hctx) && | |
965 | list_empty_careful(&hctx->dispatch)) || | |
5d12f905 | 966 | test_bit(BLK_MQ_S_STOPPED, &hctx->state)) |
320ae51f JA |
967 | continue; |
968 | ||
b94ec296 | 969 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
970 | } |
971 | } | |
b94ec296 | 972 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f JA |
973 | |
974 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) | |
975 | { | |
27489a3c | 976 | cancel_work(&hctx->run_work); |
70f4db63 | 977 | cancel_delayed_work(&hctx->delay_work); |
320ae51f JA |
978 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
979 | } | |
980 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); | |
981 | ||
280d45f6 CH |
982 | void blk_mq_stop_hw_queues(struct request_queue *q) |
983 | { | |
984 | struct blk_mq_hw_ctx *hctx; | |
985 | int i; | |
986 | ||
987 | queue_for_each_hw_ctx(q, hctx, i) | |
988 | blk_mq_stop_hw_queue(hctx); | |
989 | } | |
990 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
991 | ||
320ae51f JA |
992 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
993 | { | |
994 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 995 | |
0ffbce80 | 996 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
997 | } |
998 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
999 | ||
2f268556 CH |
1000 | void blk_mq_start_hw_queues(struct request_queue *q) |
1001 | { | |
1002 | struct blk_mq_hw_ctx *hctx; | |
1003 | int i; | |
1004 | ||
1005 | queue_for_each_hw_ctx(q, hctx, i) | |
1006 | blk_mq_start_hw_queue(hctx); | |
1007 | } | |
1008 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1009 | ||
1b4a3258 | 1010 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1011 | { |
1012 | struct blk_mq_hw_ctx *hctx; | |
1013 | int i; | |
1014 | ||
1015 | queue_for_each_hw_ctx(q, hctx, i) { | |
1016 | if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state)) | |
1017 | continue; | |
1018 | ||
1019 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1b4a3258 | 1020 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
1021 | } |
1022 | } | |
1023 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1024 | ||
70f4db63 | 1025 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
1026 | { |
1027 | struct blk_mq_hw_ctx *hctx; | |
1028 | ||
27489a3c | 1029 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work); |
e4043dcf | 1030 | |
320ae51f JA |
1031 | __blk_mq_run_hw_queue(hctx); |
1032 | } | |
1033 | ||
70f4db63 CH |
1034 | static void blk_mq_delay_work_fn(struct work_struct *work) |
1035 | { | |
1036 | struct blk_mq_hw_ctx *hctx; | |
1037 | ||
1038 | hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work); | |
1039 | ||
1040 | if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state)) | |
1041 | __blk_mq_run_hw_queue(hctx); | |
1042 | } | |
1043 | ||
1044 | void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1045 | { | |
19c66e59 ML |
1046 | if (unlikely(!blk_mq_hw_queue_mapped(hctx))) |
1047 | return; | |
70f4db63 | 1048 | |
b657d7e6 CH |
1049 | kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx), |
1050 | &hctx->delay_work, msecs_to_jiffies(msecs)); | |
70f4db63 CH |
1051 | } |
1052 | EXPORT_SYMBOL(blk_mq_delay_queue); | |
1053 | ||
cfd0c552 | 1054 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1055 | struct request *rq, |
1056 | bool at_head) | |
320ae51f | 1057 | { |
e57690fe JA |
1058 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
1059 | ||
01b983c9 JA |
1060 | trace_block_rq_insert(hctx->queue, rq); |
1061 | ||
72a0a36e CH |
1062 | if (at_head) |
1063 | list_add(&rq->queuelist, &ctx->rq_list); | |
1064 | else | |
1065 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
cfd0c552 | 1066 | } |
4bb659b1 | 1067 | |
cfd0c552 ML |
1068 | static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, |
1069 | struct request *rq, bool at_head) | |
1070 | { | |
1071 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1072 | ||
e57690fe | 1073 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1074 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1075 | } |
1076 | ||
eeabc850 | 1077 | void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue, |
e57690fe | 1078 | bool async) |
320ae51f | 1079 | { |
e57690fe | 1080 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
eeabc850 | 1081 | struct request_queue *q = rq->q; |
320ae51f | 1082 | struct blk_mq_hw_ctx *hctx; |
320ae51f | 1083 | |
320ae51f JA |
1084 | hctx = q->mq_ops->map_queue(q, ctx->cpu); |
1085 | ||
a57a178a CH |
1086 | spin_lock(&ctx->lock); |
1087 | __blk_mq_insert_request(hctx, rq, at_head); | |
1088 | spin_unlock(&ctx->lock); | |
320ae51f | 1089 | |
320ae51f JA |
1090 | if (run_queue) |
1091 | blk_mq_run_hw_queue(hctx, async); | |
1092 | } | |
1093 | ||
1094 | static void blk_mq_insert_requests(struct request_queue *q, | |
1095 | struct blk_mq_ctx *ctx, | |
1096 | struct list_head *list, | |
1097 | int depth, | |
1098 | bool from_schedule) | |
1099 | ||
1100 | { | |
1101 | struct blk_mq_hw_ctx *hctx; | |
320ae51f JA |
1102 | |
1103 | trace_block_unplug(q, depth, !from_schedule); | |
1104 | ||
320ae51f JA |
1105 | hctx = q->mq_ops->map_queue(q, ctx->cpu); |
1106 | ||
1107 | /* | |
1108 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1109 | * offline now | |
1110 | */ | |
1111 | spin_lock(&ctx->lock); | |
1112 | while (!list_empty(list)) { | |
1113 | struct request *rq; | |
1114 | ||
1115 | rq = list_first_entry(list, struct request, queuelist); | |
e57690fe | 1116 | BUG_ON(rq->mq_ctx != ctx); |
320ae51f | 1117 | list_del_init(&rq->queuelist); |
e57690fe | 1118 | __blk_mq_insert_req_list(hctx, rq, false); |
320ae51f | 1119 | } |
cfd0c552 | 1120 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1121 | spin_unlock(&ctx->lock); |
1122 | ||
320ae51f JA |
1123 | blk_mq_run_hw_queue(hctx, from_schedule); |
1124 | } | |
1125 | ||
1126 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1127 | { | |
1128 | struct request *rqa = container_of(a, struct request, queuelist); | |
1129 | struct request *rqb = container_of(b, struct request, queuelist); | |
1130 | ||
1131 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1132 | (rqa->mq_ctx == rqb->mq_ctx && | |
1133 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1134 | } | |
1135 | ||
1136 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1137 | { | |
1138 | struct blk_mq_ctx *this_ctx; | |
1139 | struct request_queue *this_q; | |
1140 | struct request *rq; | |
1141 | LIST_HEAD(list); | |
1142 | LIST_HEAD(ctx_list); | |
1143 | unsigned int depth; | |
1144 | ||
1145 | list_splice_init(&plug->mq_list, &list); | |
1146 | ||
1147 | list_sort(NULL, &list, plug_ctx_cmp); | |
1148 | ||
1149 | this_q = NULL; | |
1150 | this_ctx = NULL; | |
1151 | depth = 0; | |
1152 | ||
1153 | while (!list_empty(&list)) { | |
1154 | rq = list_entry_rq(list.next); | |
1155 | list_del_init(&rq->queuelist); | |
1156 | BUG_ON(!rq->q); | |
1157 | if (rq->mq_ctx != this_ctx) { | |
1158 | if (this_ctx) { | |
1159 | blk_mq_insert_requests(this_q, this_ctx, | |
1160 | &ctx_list, depth, | |
1161 | from_schedule); | |
1162 | } | |
1163 | ||
1164 | this_ctx = rq->mq_ctx; | |
1165 | this_q = rq->q; | |
1166 | depth = 0; | |
1167 | } | |
1168 | ||
1169 | depth++; | |
1170 | list_add_tail(&rq->queuelist, &ctx_list); | |
1171 | } | |
1172 | ||
1173 | /* | |
1174 | * If 'this_ctx' is set, we know we have entries to complete | |
1175 | * on 'ctx_list'. Do those. | |
1176 | */ | |
1177 | if (this_ctx) { | |
1178 | blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth, | |
1179 | from_schedule); | |
1180 | } | |
1181 | } | |
1182 | ||
1183 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1184 | { | |
1185 | init_request_from_bio(rq, bio); | |
4b570521 | 1186 | |
a21f2a3e | 1187 | blk_account_io_start(rq, 1); |
320ae51f JA |
1188 | } |
1189 | ||
274a5843 JA |
1190 | static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx) |
1191 | { | |
1192 | return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) && | |
1193 | !blk_queue_nomerges(hctx->queue); | |
1194 | } | |
1195 | ||
07068d5b JA |
1196 | static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx, |
1197 | struct blk_mq_ctx *ctx, | |
1198 | struct request *rq, struct bio *bio) | |
320ae51f | 1199 | { |
e18378a6 | 1200 | if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) { |
07068d5b JA |
1201 | blk_mq_bio_to_request(rq, bio); |
1202 | spin_lock(&ctx->lock); | |
1203 | insert_rq: | |
1204 | __blk_mq_insert_request(hctx, rq, false); | |
1205 | spin_unlock(&ctx->lock); | |
1206 | return false; | |
1207 | } else { | |
274a5843 JA |
1208 | struct request_queue *q = hctx->queue; |
1209 | ||
07068d5b JA |
1210 | spin_lock(&ctx->lock); |
1211 | if (!blk_mq_attempt_merge(q, ctx, bio)) { | |
1212 | blk_mq_bio_to_request(rq, bio); | |
1213 | goto insert_rq; | |
1214 | } | |
320ae51f | 1215 | |
07068d5b JA |
1216 | spin_unlock(&ctx->lock); |
1217 | __blk_mq_free_request(hctx, ctx, rq); | |
1218 | return true; | |
14ec77f3 | 1219 | } |
07068d5b | 1220 | } |
14ec77f3 | 1221 | |
07068d5b JA |
1222 | struct blk_map_ctx { |
1223 | struct blk_mq_hw_ctx *hctx; | |
1224 | struct blk_mq_ctx *ctx; | |
1225 | }; | |
1226 | ||
1227 | static struct request *blk_mq_map_request(struct request_queue *q, | |
1228 | struct bio *bio, | |
1229 | struct blk_map_ctx *data) | |
1230 | { | |
1231 | struct blk_mq_hw_ctx *hctx; | |
1232 | struct blk_mq_ctx *ctx; | |
1233 | struct request *rq; | |
cc6e3b10 MC |
1234 | int op = bio_data_dir(bio); |
1235 | int op_flags = 0; | |
cb96a42c | 1236 | struct blk_mq_alloc_data alloc_data; |
320ae51f | 1237 | |
3ef28e83 | 1238 | blk_queue_enter_live(q); |
320ae51f JA |
1239 | ctx = blk_mq_get_ctx(q); |
1240 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
1241 | ||
1eff9d32 | 1242 | if (rw_is_sync(bio_op(bio), bio->bi_opf)) |
cc6e3b10 | 1243 | op_flags |= REQ_SYNC; |
07068d5b | 1244 | |
cc6e3b10 | 1245 | trace_block_getrq(q, bio, op); |
6f3b0e8b | 1246 | blk_mq_set_alloc_data(&alloc_data, q, BLK_MQ_REQ_NOWAIT, ctx, hctx); |
cc6e3b10 | 1247 | rq = __blk_mq_alloc_request(&alloc_data, op, op_flags); |
5dee8577 | 1248 | if (unlikely(!rq)) { |
793597a6 | 1249 | __blk_mq_run_hw_queue(hctx); |
320ae51f | 1250 | blk_mq_put_ctx(ctx); |
cc6e3b10 | 1251 | trace_block_sleeprq(q, bio, op); |
793597a6 CH |
1252 | |
1253 | ctx = blk_mq_get_ctx(q); | |
320ae51f | 1254 | hctx = q->mq_ops->map_queue(q, ctx->cpu); |
6f3b0e8b | 1255 | blk_mq_set_alloc_data(&alloc_data, q, 0, ctx, hctx); |
cc6e3b10 | 1256 | rq = __blk_mq_alloc_request(&alloc_data, op, op_flags); |
cb96a42c ML |
1257 | ctx = alloc_data.ctx; |
1258 | hctx = alloc_data.hctx; | |
320ae51f JA |
1259 | } |
1260 | ||
1261 | hctx->queued++; | |
07068d5b JA |
1262 | data->hctx = hctx; |
1263 | data->ctx = ctx; | |
1264 | return rq; | |
1265 | } | |
1266 | ||
7b371636 | 1267 | static int blk_mq_direct_issue_request(struct request *rq, blk_qc_t *cookie) |
f984df1f SL |
1268 | { |
1269 | int ret; | |
1270 | struct request_queue *q = rq->q; | |
1271 | struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q, | |
1272 | rq->mq_ctx->cpu); | |
1273 | struct blk_mq_queue_data bd = { | |
1274 | .rq = rq, | |
1275 | .list = NULL, | |
1276 | .last = 1 | |
1277 | }; | |
7b371636 | 1278 | blk_qc_t new_cookie = blk_tag_to_qc_t(rq->tag, hctx->queue_num); |
f984df1f SL |
1279 | |
1280 | /* | |
1281 | * For OK queue, we are done. For error, kill it. Any other | |
1282 | * error (busy), just add it to our list as we previously | |
1283 | * would have done | |
1284 | */ | |
1285 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
7b371636 JA |
1286 | if (ret == BLK_MQ_RQ_QUEUE_OK) { |
1287 | *cookie = new_cookie; | |
f984df1f | 1288 | return 0; |
7b371636 | 1289 | } |
f984df1f | 1290 | |
7b371636 JA |
1291 | __blk_mq_requeue_request(rq); |
1292 | ||
1293 | if (ret == BLK_MQ_RQ_QUEUE_ERROR) { | |
1294 | *cookie = BLK_QC_T_NONE; | |
1295 | rq->errors = -EIO; | |
1296 | blk_mq_end_request(rq, rq->errors); | |
1297 | return 0; | |
f984df1f | 1298 | } |
7b371636 JA |
1299 | |
1300 | return -1; | |
f984df1f SL |
1301 | } |
1302 | ||
07068d5b JA |
1303 | /* |
1304 | * Multiple hardware queue variant. This will not use per-process plugs, | |
1305 | * but will attempt to bypass the hctx queueing if we can go straight to | |
1306 | * hardware for SYNC IO. | |
1307 | */ | |
dece1635 | 1308 | static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1309 | { |
1eff9d32 JA |
1310 | const int is_sync = rw_is_sync(bio_op(bio), bio->bi_opf); |
1311 | const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA); | |
07068d5b JA |
1312 | struct blk_map_ctx data; |
1313 | struct request *rq; | |
f984df1f SL |
1314 | unsigned int request_count = 0; |
1315 | struct blk_plug *plug; | |
5b3f341f | 1316 | struct request *same_queue_rq = NULL; |
7b371636 | 1317 | blk_qc_t cookie; |
07068d5b JA |
1318 | |
1319 | blk_queue_bounce(q, &bio); | |
1320 | ||
1321 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
4246a0b6 | 1322 | bio_io_error(bio); |
dece1635 | 1323 | return BLK_QC_T_NONE; |
07068d5b JA |
1324 | } |
1325 | ||
54efd50b KO |
1326 | blk_queue_split(q, &bio, q->bio_split); |
1327 | ||
87c279e6 OS |
1328 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
1329 | blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq)) | |
1330 | return BLK_QC_T_NONE; | |
f984df1f | 1331 | |
07068d5b JA |
1332 | rq = blk_mq_map_request(q, bio, &data); |
1333 | if (unlikely(!rq)) | |
dece1635 | 1334 | return BLK_QC_T_NONE; |
07068d5b | 1335 | |
7b371636 | 1336 | cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num); |
07068d5b JA |
1337 | |
1338 | if (unlikely(is_flush_fua)) { | |
1339 | blk_mq_bio_to_request(rq, bio); | |
1340 | blk_insert_flush(rq); | |
1341 | goto run_queue; | |
1342 | } | |
1343 | ||
f984df1f | 1344 | plug = current->plug; |
e167dfb5 JA |
1345 | /* |
1346 | * If the driver supports defer issued based on 'last', then | |
1347 | * queue it up like normal since we can potentially save some | |
1348 | * CPU this way. | |
1349 | */ | |
f984df1f SL |
1350 | if (((plug && !blk_queue_nomerges(q)) || is_sync) && |
1351 | !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) { | |
1352 | struct request *old_rq = NULL; | |
07068d5b JA |
1353 | |
1354 | blk_mq_bio_to_request(rq, bio); | |
07068d5b JA |
1355 | |
1356 | /* | |
b094f89c | 1357 | * We do limited pluging. If the bio can be merged, do that. |
f984df1f SL |
1358 | * Otherwise the existing request in the plug list will be |
1359 | * issued. So the plug list will have one request at most | |
07068d5b | 1360 | */ |
f984df1f | 1361 | if (plug) { |
5b3f341f SL |
1362 | /* |
1363 | * The plug list might get flushed before this. If that | |
b094f89c JA |
1364 | * happens, same_queue_rq is invalid and plug list is |
1365 | * empty | |
1366 | */ | |
5b3f341f SL |
1367 | if (same_queue_rq && !list_empty(&plug->mq_list)) { |
1368 | old_rq = same_queue_rq; | |
f984df1f | 1369 | list_del_init(&old_rq->queuelist); |
07068d5b | 1370 | } |
f984df1f SL |
1371 | list_add_tail(&rq->queuelist, &plug->mq_list); |
1372 | } else /* is_sync */ | |
1373 | old_rq = rq; | |
1374 | blk_mq_put_ctx(data.ctx); | |
1375 | if (!old_rq) | |
7b371636 JA |
1376 | goto done; |
1377 | if (!blk_mq_direct_issue_request(old_rq, &cookie)) | |
1378 | goto done; | |
f984df1f | 1379 | blk_mq_insert_request(old_rq, false, true, true); |
7b371636 | 1380 | goto done; |
07068d5b JA |
1381 | } |
1382 | ||
1383 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { | |
1384 | /* | |
1385 | * For a SYNC request, send it to the hardware immediately. For | |
1386 | * an ASYNC request, just ensure that we run it later on. The | |
1387 | * latter allows for merging opportunities and more efficient | |
1388 | * dispatching. | |
1389 | */ | |
1390 | run_queue: | |
1391 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); | |
1392 | } | |
07068d5b | 1393 | blk_mq_put_ctx(data.ctx); |
7b371636 JA |
1394 | done: |
1395 | return cookie; | |
07068d5b JA |
1396 | } |
1397 | ||
1398 | /* | |
1399 | * Single hardware queue variant. This will attempt to use any per-process | |
1400 | * plug for merging and IO deferral. | |
1401 | */ | |
dece1635 | 1402 | static blk_qc_t blk_sq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1403 | { |
1eff9d32 JA |
1404 | const int is_sync = rw_is_sync(bio_op(bio), bio->bi_opf); |
1405 | const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA); | |
e6c4438b JM |
1406 | struct blk_plug *plug; |
1407 | unsigned int request_count = 0; | |
07068d5b JA |
1408 | struct blk_map_ctx data; |
1409 | struct request *rq; | |
7b371636 | 1410 | blk_qc_t cookie; |
07068d5b | 1411 | |
07068d5b JA |
1412 | blk_queue_bounce(q, &bio); |
1413 | ||
1414 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
4246a0b6 | 1415 | bio_io_error(bio); |
dece1635 | 1416 | return BLK_QC_T_NONE; |
07068d5b JA |
1417 | } |
1418 | ||
54efd50b KO |
1419 | blk_queue_split(q, &bio, q->bio_split); |
1420 | ||
87c279e6 OS |
1421 | if (!is_flush_fua && !blk_queue_nomerges(q)) { |
1422 | if (blk_attempt_plug_merge(q, bio, &request_count, NULL)) | |
1423 | return BLK_QC_T_NONE; | |
1424 | } else | |
1425 | request_count = blk_plug_queued_count(q); | |
07068d5b JA |
1426 | |
1427 | rq = blk_mq_map_request(q, bio, &data); | |
ff87bcec | 1428 | if (unlikely(!rq)) |
dece1635 | 1429 | return BLK_QC_T_NONE; |
320ae51f | 1430 | |
7b371636 | 1431 | cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num); |
320ae51f JA |
1432 | |
1433 | if (unlikely(is_flush_fua)) { | |
1434 | blk_mq_bio_to_request(rq, bio); | |
320ae51f JA |
1435 | blk_insert_flush(rq); |
1436 | goto run_queue; | |
1437 | } | |
1438 | ||
1439 | /* | |
1440 | * A task plug currently exists. Since this is completely lockless, | |
1441 | * utilize that to temporarily store requests until the task is | |
1442 | * either done or scheduled away. | |
1443 | */ | |
e6c4438b JM |
1444 | plug = current->plug; |
1445 | if (plug) { | |
1446 | blk_mq_bio_to_request(rq, bio); | |
676d0607 | 1447 | if (!request_count) |
e6c4438b | 1448 | trace_block_plug(q); |
b094f89c JA |
1449 | |
1450 | blk_mq_put_ctx(data.ctx); | |
1451 | ||
1452 | if (request_count >= BLK_MAX_REQUEST_COUNT) { | |
e6c4438b JM |
1453 | blk_flush_plug_list(plug, false); |
1454 | trace_block_plug(q); | |
320ae51f | 1455 | } |
b094f89c | 1456 | |
e6c4438b | 1457 | list_add_tail(&rq->queuelist, &plug->mq_list); |
7b371636 | 1458 | return cookie; |
320ae51f JA |
1459 | } |
1460 | ||
07068d5b JA |
1461 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { |
1462 | /* | |
1463 | * For a SYNC request, send it to the hardware immediately. For | |
1464 | * an ASYNC request, just ensure that we run it later on. The | |
1465 | * latter allows for merging opportunities and more efficient | |
1466 | * dispatching. | |
1467 | */ | |
1468 | run_queue: | |
1469 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); | |
320ae51f JA |
1470 | } |
1471 | ||
07068d5b | 1472 | blk_mq_put_ctx(data.ctx); |
7b371636 | 1473 | return cookie; |
320ae51f JA |
1474 | } |
1475 | ||
1476 | /* | |
1477 | * Default mapping to a software queue, since we use one per CPU. | |
1478 | */ | |
1479 | struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu) | |
1480 | { | |
1481 | return q->queue_hw_ctx[q->mq_map[cpu]]; | |
1482 | } | |
1483 | EXPORT_SYMBOL(blk_mq_map_queue); | |
1484 | ||
24d2f903 CH |
1485 | static void blk_mq_free_rq_map(struct blk_mq_tag_set *set, |
1486 | struct blk_mq_tags *tags, unsigned int hctx_idx) | |
95363efd | 1487 | { |
e9b267d9 | 1488 | struct page *page; |
320ae51f | 1489 | |
24d2f903 | 1490 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 1491 | int i; |
320ae51f | 1492 | |
24d2f903 CH |
1493 | for (i = 0; i < tags->nr_tags; i++) { |
1494 | if (!tags->rqs[i]) | |
e9b267d9 | 1495 | continue; |
24d2f903 CH |
1496 | set->ops->exit_request(set->driver_data, tags->rqs[i], |
1497 | hctx_idx, i); | |
a5164405 | 1498 | tags->rqs[i] = NULL; |
e9b267d9 | 1499 | } |
320ae51f | 1500 | } |
320ae51f | 1501 | |
24d2f903 CH |
1502 | while (!list_empty(&tags->page_list)) { |
1503 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 1504 | list_del_init(&page->lru); |
f75782e4 CM |
1505 | /* |
1506 | * Remove kmemleak object previously allocated in | |
1507 | * blk_mq_init_rq_map(). | |
1508 | */ | |
1509 | kmemleak_free(page_address(page)); | |
320ae51f JA |
1510 | __free_pages(page, page->private); |
1511 | } | |
1512 | ||
24d2f903 | 1513 | kfree(tags->rqs); |
320ae51f | 1514 | |
24d2f903 | 1515 | blk_mq_free_tags(tags); |
320ae51f JA |
1516 | } |
1517 | ||
1518 | static size_t order_to_size(unsigned int order) | |
1519 | { | |
4ca08500 | 1520 | return (size_t)PAGE_SIZE << order; |
320ae51f JA |
1521 | } |
1522 | ||
24d2f903 CH |
1523 | static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set, |
1524 | unsigned int hctx_idx) | |
320ae51f | 1525 | { |
24d2f903 | 1526 | struct blk_mq_tags *tags; |
320ae51f JA |
1527 | unsigned int i, j, entries_per_page, max_order = 4; |
1528 | size_t rq_size, left; | |
1529 | ||
24d2f903 | 1530 | tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags, |
24391c0d SL |
1531 | set->numa_node, |
1532 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); | |
24d2f903 CH |
1533 | if (!tags) |
1534 | return NULL; | |
320ae51f | 1535 | |
24d2f903 CH |
1536 | INIT_LIST_HEAD(&tags->page_list); |
1537 | ||
a5164405 JA |
1538 | tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *), |
1539 | GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY, | |
1540 | set->numa_node); | |
24d2f903 CH |
1541 | if (!tags->rqs) { |
1542 | blk_mq_free_tags(tags); | |
1543 | return NULL; | |
1544 | } | |
320ae51f JA |
1545 | |
1546 | /* | |
1547 | * rq_size is the size of the request plus driver payload, rounded | |
1548 | * to the cacheline size | |
1549 | */ | |
24d2f903 | 1550 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 1551 | cache_line_size()); |
24d2f903 | 1552 | left = rq_size * set->queue_depth; |
320ae51f | 1553 | |
24d2f903 | 1554 | for (i = 0; i < set->queue_depth; ) { |
320ae51f JA |
1555 | int this_order = max_order; |
1556 | struct page *page; | |
1557 | int to_do; | |
1558 | void *p; | |
1559 | ||
b3a834b1 | 1560 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
1561 | this_order--; |
1562 | ||
1563 | do { | |
a5164405 | 1564 | page = alloc_pages_node(set->numa_node, |
ac211175 | 1565 | GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 1566 | this_order); |
320ae51f JA |
1567 | if (page) |
1568 | break; | |
1569 | if (!this_order--) | |
1570 | break; | |
1571 | if (order_to_size(this_order) < rq_size) | |
1572 | break; | |
1573 | } while (1); | |
1574 | ||
1575 | if (!page) | |
24d2f903 | 1576 | goto fail; |
320ae51f JA |
1577 | |
1578 | page->private = this_order; | |
24d2f903 | 1579 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
1580 | |
1581 | p = page_address(page); | |
f75782e4 CM |
1582 | /* |
1583 | * Allow kmemleak to scan these pages as they contain pointers | |
1584 | * to additional allocations like via ops->init_request(). | |
1585 | */ | |
1586 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_KERNEL); | |
320ae51f | 1587 | entries_per_page = order_to_size(this_order) / rq_size; |
24d2f903 | 1588 | to_do = min(entries_per_page, set->queue_depth - i); |
320ae51f JA |
1589 | left -= to_do * rq_size; |
1590 | for (j = 0; j < to_do; j++) { | |
24d2f903 CH |
1591 | tags->rqs[i] = p; |
1592 | if (set->ops->init_request) { | |
1593 | if (set->ops->init_request(set->driver_data, | |
1594 | tags->rqs[i], hctx_idx, i, | |
a5164405 JA |
1595 | set->numa_node)) { |
1596 | tags->rqs[i] = NULL; | |
24d2f903 | 1597 | goto fail; |
a5164405 | 1598 | } |
e9b267d9 CH |
1599 | } |
1600 | ||
320ae51f JA |
1601 | p += rq_size; |
1602 | i++; | |
1603 | } | |
1604 | } | |
24d2f903 | 1605 | return tags; |
320ae51f | 1606 | |
24d2f903 | 1607 | fail: |
24d2f903 CH |
1608 | blk_mq_free_rq_map(set, tags, hctx_idx); |
1609 | return NULL; | |
320ae51f JA |
1610 | } |
1611 | ||
1429d7c9 JA |
1612 | static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap) |
1613 | { | |
1614 | kfree(bitmap->map); | |
1615 | } | |
1616 | ||
1617 | static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node) | |
1618 | { | |
1619 | unsigned int bpw = 8, total, num_maps, i; | |
1620 | ||
1621 | bitmap->bits_per_word = bpw; | |
1622 | ||
1623 | num_maps = ALIGN(nr_cpu_ids, bpw) / bpw; | |
1624 | bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap), | |
1625 | GFP_KERNEL, node); | |
1626 | if (!bitmap->map) | |
1627 | return -ENOMEM; | |
1628 | ||
1429d7c9 JA |
1629 | total = nr_cpu_ids; |
1630 | for (i = 0; i < num_maps; i++) { | |
1631 | bitmap->map[i].depth = min(total, bitmap->bits_per_word); | |
1632 | total -= bitmap->map[i].depth; | |
1633 | } | |
1634 | ||
1635 | return 0; | |
1636 | } | |
1637 | ||
e57690fe JA |
1638 | /* |
1639 | * 'cpu' is going away. splice any existing rq_list entries from this | |
1640 | * software queue to the hw queue dispatch list, and ensure that it | |
1641 | * gets run. | |
1642 | */ | |
484b4061 JA |
1643 | static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu) |
1644 | { | |
484b4061 JA |
1645 | struct blk_mq_ctx *ctx; |
1646 | LIST_HEAD(tmp); | |
1647 | ||
e57690fe | 1648 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
484b4061 JA |
1649 | |
1650 | spin_lock(&ctx->lock); | |
1651 | if (!list_empty(&ctx->rq_list)) { | |
1652 | list_splice_init(&ctx->rq_list, &tmp); | |
1653 | blk_mq_hctx_clear_pending(hctx, ctx); | |
1654 | } | |
1655 | spin_unlock(&ctx->lock); | |
1656 | ||
1657 | if (list_empty(&tmp)) | |
1658 | return NOTIFY_OK; | |
1659 | ||
e57690fe JA |
1660 | spin_lock(&hctx->lock); |
1661 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
1662 | spin_unlock(&hctx->lock); | |
484b4061 JA |
1663 | |
1664 | blk_mq_run_hw_queue(hctx, true); | |
484b4061 JA |
1665 | return NOTIFY_OK; |
1666 | } | |
1667 | ||
484b4061 JA |
1668 | static int blk_mq_hctx_notify(void *data, unsigned long action, |
1669 | unsigned int cpu) | |
1670 | { | |
1671 | struct blk_mq_hw_ctx *hctx = data; | |
1672 | ||
1673 | if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) | |
1674 | return blk_mq_hctx_cpu_offline(hctx, cpu); | |
2a34c087 ML |
1675 | |
1676 | /* | |
1677 | * In case of CPU online, tags may be reallocated | |
1678 | * in blk_mq_map_swqueue() after mapping is updated. | |
1679 | */ | |
484b4061 JA |
1680 | |
1681 | return NOTIFY_OK; | |
1682 | } | |
1683 | ||
c3b4afca | 1684 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
1685 | static void blk_mq_exit_hctx(struct request_queue *q, |
1686 | struct blk_mq_tag_set *set, | |
1687 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
1688 | { | |
f70ced09 ML |
1689 | unsigned flush_start_tag = set->queue_depth; |
1690 | ||
08e98fc6 ML |
1691 | blk_mq_tag_idle(hctx); |
1692 | ||
f70ced09 ML |
1693 | if (set->ops->exit_request) |
1694 | set->ops->exit_request(set->driver_data, | |
1695 | hctx->fq->flush_rq, hctx_idx, | |
1696 | flush_start_tag + hctx_idx); | |
1697 | ||
08e98fc6 ML |
1698 | if (set->ops->exit_hctx) |
1699 | set->ops->exit_hctx(hctx, hctx_idx); | |
1700 | ||
1701 | blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier); | |
f70ced09 | 1702 | blk_free_flush_queue(hctx->fq); |
08e98fc6 ML |
1703 | blk_mq_free_bitmap(&hctx->ctx_map); |
1704 | } | |
1705 | ||
624dbe47 ML |
1706 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
1707 | struct blk_mq_tag_set *set, int nr_queue) | |
1708 | { | |
1709 | struct blk_mq_hw_ctx *hctx; | |
1710 | unsigned int i; | |
1711 | ||
1712 | queue_for_each_hw_ctx(q, hctx, i) { | |
1713 | if (i == nr_queue) | |
1714 | break; | |
08e98fc6 | 1715 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 1716 | } |
624dbe47 ML |
1717 | } |
1718 | ||
1719 | static void blk_mq_free_hw_queues(struct request_queue *q, | |
1720 | struct blk_mq_tag_set *set) | |
1721 | { | |
1722 | struct blk_mq_hw_ctx *hctx; | |
1723 | unsigned int i; | |
1724 | ||
e09aae7e | 1725 | queue_for_each_hw_ctx(q, hctx, i) |
624dbe47 | 1726 | free_cpumask_var(hctx->cpumask); |
624dbe47 ML |
1727 | } |
1728 | ||
08e98fc6 ML |
1729 | static int blk_mq_init_hctx(struct request_queue *q, |
1730 | struct blk_mq_tag_set *set, | |
1731 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 1732 | { |
08e98fc6 | 1733 | int node; |
f70ced09 | 1734 | unsigned flush_start_tag = set->queue_depth; |
08e98fc6 ML |
1735 | |
1736 | node = hctx->numa_node; | |
1737 | if (node == NUMA_NO_NODE) | |
1738 | node = hctx->numa_node = set->numa_node; | |
1739 | ||
27489a3c | 1740 | INIT_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
1741 | INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn); |
1742 | spin_lock_init(&hctx->lock); | |
1743 | INIT_LIST_HEAD(&hctx->dispatch); | |
1744 | hctx->queue = q; | |
1745 | hctx->queue_num = hctx_idx; | |
2404e607 | 1746 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; |
08e98fc6 ML |
1747 | |
1748 | blk_mq_init_cpu_notifier(&hctx->cpu_notifier, | |
1749 | blk_mq_hctx_notify, hctx); | |
1750 | blk_mq_register_cpu_notifier(&hctx->cpu_notifier); | |
1751 | ||
1752 | hctx->tags = set->tags[hctx_idx]; | |
320ae51f JA |
1753 | |
1754 | /* | |
08e98fc6 ML |
1755 | * Allocate space for all possible cpus to avoid allocation at |
1756 | * runtime | |
320ae51f | 1757 | */ |
08e98fc6 ML |
1758 | hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *), |
1759 | GFP_KERNEL, node); | |
1760 | if (!hctx->ctxs) | |
1761 | goto unregister_cpu_notifier; | |
320ae51f | 1762 | |
08e98fc6 ML |
1763 | if (blk_mq_alloc_bitmap(&hctx->ctx_map, node)) |
1764 | goto free_ctxs; | |
320ae51f | 1765 | |
08e98fc6 | 1766 | hctx->nr_ctx = 0; |
320ae51f | 1767 | |
08e98fc6 ML |
1768 | if (set->ops->init_hctx && |
1769 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
1770 | goto free_bitmap; | |
320ae51f | 1771 | |
f70ced09 ML |
1772 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size); |
1773 | if (!hctx->fq) | |
1774 | goto exit_hctx; | |
320ae51f | 1775 | |
f70ced09 ML |
1776 | if (set->ops->init_request && |
1777 | set->ops->init_request(set->driver_data, | |
1778 | hctx->fq->flush_rq, hctx_idx, | |
1779 | flush_start_tag + hctx_idx, node)) | |
1780 | goto free_fq; | |
320ae51f | 1781 | |
08e98fc6 | 1782 | return 0; |
320ae51f | 1783 | |
f70ced09 ML |
1784 | free_fq: |
1785 | kfree(hctx->fq); | |
1786 | exit_hctx: | |
1787 | if (set->ops->exit_hctx) | |
1788 | set->ops->exit_hctx(hctx, hctx_idx); | |
08e98fc6 ML |
1789 | free_bitmap: |
1790 | blk_mq_free_bitmap(&hctx->ctx_map); | |
1791 | free_ctxs: | |
1792 | kfree(hctx->ctxs); | |
1793 | unregister_cpu_notifier: | |
1794 | blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier); | |
320ae51f | 1795 | |
08e98fc6 ML |
1796 | return -1; |
1797 | } | |
320ae51f | 1798 | |
320ae51f JA |
1799 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
1800 | unsigned int nr_hw_queues) | |
1801 | { | |
1802 | unsigned int i; | |
1803 | ||
1804 | for_each_possible_cpu(i) { | |
1805 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
1806 | struct blk_mq_hw_ctx *hctx; | |
1807 | ||
1808 | memset(__ctx, 0, sizeof(*__ctx)); | |
1809 | __ctx->cpu = i; | |
1810 | spin_lock_init(&__ctx->lock); | |
1811 | INIT_LIST_HEAD(&__ctx->rq_list); | |
1812 | __ctx->queue = q; | |
1813 | ||
1814 | /* If the cpu isn't online, the cpu is mapped to first hctx */ | |
320ae51f JA |
1815 | if (!cpu_online(i)) |
1816 | continue; | |
1817 | ||
e4043dcf | 1818 | hctx = q->mq_ops->map_queue(q, i); |
e4043dcf | 1819 | |
320ae51f JA |
1820 | /* |
1821 | * Set local node, IFF we have more than one hw queue. If | |
1822 | * not, we remain on the home node of the device | |
1823 | */ | |
1824 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
bffed457 | 1825 | hctx->numa_node = local_memory_node(cpu_to_node(i)); |
320ae51f JA |
1826 | } |
1827 | } | |
1828 | ||
5778322e AM |
1829 | static void blk_mq_map_swqueue(struct request_queue *q, |
1830 | const struct cpumask *online_mask) | |
320ae51f JA |
1831 | { |
1832 | unsigned int i; | |
1833 | struct blk_mq_hw_ctx *hctx; | |
1834 | struct blk_mq_ctx *ctx; | |
2a34c087 | 1835 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 1836 | |
60de074b AM |
1837 | /* |
1838 | * Avoid others reading imcomplete hctx->cpumask through sysfs | |
1839 | */ | |
1840 | mutex_lock(&q->sysfs_lock); | |
1841 | ||
320ae51f | 1842 | queue_for_each_hw_ctx(q, hctx, i) { |
e4043dcf | 1843 | cpumask_clear(hctx->cpumask); |
320ae51f JA |
1844 | hctx->nr_ctx = 0; |
1845 | } | |
1846 | ||
1847 | /* | |
1848 | * Map software to hardware queues | |
1849 | */ | |
897bb0c7 | 1850 | for_each_possible_cpu(i) { |
320ae51f | 1851 | /* If the cpu isn't online, the cpu is mapped to first hctx */ |
5778322e | 1852 | if (!cpumask_test_cpu(i, online_mask)) |
e4043dcf JA |
1853 | continue; |
1854 | ||
897bb0c7 | 1855 | ctx = per_cpu_ptr(q->queue_ctx, i); |
320ae51f | 1856 | hctx = q->mq_ops->map_queue(q, i); |
868f2f0b | 1857 | |
e4043dcf | 1858 | cpumask_set_cpu(i, hctx->cpumask); |
320ae51f JA |
1859 | ctx->index_hw = hctx->nr_ctx; |
1860 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
1861 | } | |
506e931f | 1862 | |
60de074b AM |
1863 | mutex_unlock(&q->sysfs_lock); |
1864 | ||
506e931f | 1865 | queue_for_each_hw_ctx(q, hctx, i) { |
889fa31f CY |
1866 | struct blk_mq_ctxmap *map = &hctx->ctx_map; |
1867 | ||
484b4061 | 1868 | /* |
a68aafa5 JA |
1869 | * If no software queues are mapped to this hardware queue, |
1870 | * disable it and free the request entries. | |
484b4061 JA |
1871 | */ |
1872 | if (!hctx->nr_ctx) { | |
484b4061 JA |
1873 | if (set->tags[i]) { |
1874 | blk_mq_free_rq_map(set, set->tags[i], i); | |
1875 | set->tags[i] = NULL; | |
484b4061 | 1876 | } |
2a34c087 | 1877 | hctx->tags = NULL; |
484b4061 JA |
1878 | continue; |
1879 | } | |
1880 | ||
2a34c087 ML |
1881 | /* unmapped hw queue can be remapped after CPU topo changed */ |
1882 | if (!set->tags[i]) | |
1883 | set->tags[i] = blk_mq_init_rq_map(set, i); | |
1884 | hctx->tags = set->tags[i]; | |
1885 | WARN_ON(!hctx->tags); | |
1886 | ||
e0e827b9 | 1887 | cpumask_copy(hctx->tags->cpumask, hctx->cpumask); |
889fa31f CY |
1888 | /* |
1889 | * Set the map size to the number of mapped software queues. | |
1890 | * This is more accurate and more efficient than looping | |
1891 | * over all possibly mapped software queues. | |
1892 | */ | |
569fd0ce | 1893 | map->size = DIV_ROUND_UP(hctx->nr_ctx, map->bits_per_word); |
889fa31f | 1894 | |
484b4061 JA |
1895 | /* |
1896 | * Initialize batch roundrobin counts | |
1897 | */ | |
506e931f JA |
1898 | hctx->next_cpu = cpumask_first(hctx->cpumask); |
1899 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
1900 | } | |
320ae51f JA |
1901 | } |
1902 | ||
2404e607 | 1903 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
1904 | { |
1905 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
1906 | int i; |
1907 | ||
2404e607 JM |
1908 | queue_for_each_hw_ctx(q, hctx, i) { |
1909 | if (shared) | |
1910 | hctx->flags |= BLK_MQ_F_TAG_SHARED; | |
1911 | else | |
1912 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; | |
1913 | } | |
1914 | } | |
1915 | ||
1916 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared) | |
1917 | { | |
1918 | struct request_queue *q; | |
0d2602ca JA |
1919 | |
1920 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
1921 | blk_mq_freeze_queue(q); | |
2404e607 | 1922 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
1923 | blk_mq_unfreeze_queue(q); |
1924 | } | |
1925 | } | |
1926 | ||
1927 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
1928 | { | |
1929 | struct blk_mq_tag_set *set = q->tag_set; | |
1930 | ||
0d2602ca JA |
1931 | mutex_lock(&set->tag_list_lock); |
1932 | list_del_init(&q->tag_set_list); | |
2404e607 JM |
1933 | if (list_is_singular(&set->tag_list)) { |
1934 | /* just transitioned to unshared */ | |
1935 | set->flags &= ~BLK_MQ_F_TAG_SHARED; | |
1936 | /* update existing queue */ | |
1937 | blk_mq_update_tag_set_depth(set, false); | |
1938 | } | |
0d2602ca | 1939 | mutex_unlock(&set->tag_list_lock); |
0d2602ca JA |
1940 | } |
1941 | ||
1942 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
1943 | struct request_queue *q) | |
1944 | { | |
1945 | q->tag_set = set; | |
1946 | ||
1947 | mutex_lock(&set->tag_list_lock); | |
2404e607 JM |
1948 | |
1949 | /* Check to see if we're transitioning to shared (from 1 to 2 queues). */ | |
1950 | if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) { | |
1951 | set->flags |= BLK_MQ_F_TAG_SHARED; | |
1952 | /* update existing queue */ | |
1953 | blk_mq_update_tag_set_depth(set, true); | |
1954 | } | |
1955 | if (set->flags & BLK_MQ_F_TAG_SHARED) | |
1956 | queue_set_hctx_shared(q, true); | |
0d2602ca | 1957 | list_add_tail(&q->tag_set_list, &set->tag_list); |
2404e607 | 1958 | |
0d2602ca JA |
1959 | mutex_unlock(&set->tag_list_lock); |
1960 | } | |
1961 | ||
e09aae7e ML |
1962 | /* |
1963 | * It is the actual release handler for mq, but we do it from | |
1964 | * request queue's release handler for avoiding use-after-free | |
1965 | * and headache because q->mq_kobj shouldn't have been introduced, | |
1966 | * but we can't group ctx/kctx kobj without it. | |
1967 | */ | |
1968 | void blk_mq_release(struct request_queue *q) | |
1969 | { | |
1970 | struct blk_mq_hw_ctx *hctx; | |
1971 | unsigned int i; | |
1972 | ||
1973 | /* hctx kobj stays in hctx */ | |
c3b4afca ML |
1974 | queue_for_each_hw_ctx(q, hctx, i) { |
1975 | if (!hctx) | |
1976 | continue; | |
1977 | kfree(hctx->ctxs); | |
e09aae7e | 1978 | kfree(hctx); |
c3b4afca | 1979 | } |
e09aae7e | 1980 | |
a723bab3 AM |
1981 | kfree(q->mq_map); |
1982 | q->mq_map = NULL; | |
1983 | ||
e09aae7e ML |
1984 | kfree(q->queue_hw_ctx); |
1985 | ||
1986 | /* ctx kobj stays in queue_ctx */ | |
1987 | free_percpu(q->queue_ctx); | |
1988 | } | |
1989 | ||
24d2f903 | 1990 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
b62c21b7 MS |
1991 | { |
1992 | struct request_queue *uninit_q, *q; | |
1993 | ||
1994 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); | |
1995 | if (!uninit_q) | |
1996 | return ERR_PTR(-ENOMEM); | |
1997 | ||
1998 | q = blk_mq_init_allocated_queue(set, uninit_q); | |
1999 | if (IS_ERR(q)) | |
2000 | blk_cleanup_queue(uninit_q); | |
2001 | ||
2002 | return q; | |
2003 | } | |
2004 | EXPORT_SYMBOL(blk_mq_init_queue); | |
2005 | ||
868f2f0b KB |
2006 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
2007 | struct request_queue *q) | |
320ae51f | 2008 | { |
868f2f0b KB |
2009 | int i, j; |
2010 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; | |
f14bbe77 | 2011 | |
868f2f0b | 2012 | blk_mq_sysfs_unregister(q); |
24d2f903 | 2013 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 2014 | int node; |
f14bbe77 | 2015 | |
868f2f0b KB |
2016 | if (hctxs[i]) |
2017 | continue; | |
2018 | ||
2019 | node = blk_mq_hw_queue_to_node(q->mq_map, i); | |
cdef54dd CH |
2020 | hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx), |
2021 | GFP_KERNEL, node); | |
320ae51f | 2022 | if (!hctxs[i]) |
868f2f0b | 2023 | break; |
320ae51f | 2024 | |
a86073e4 | 2025 | if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL, |
868f2f0b KB |
2026 | node)) { |
2027 | kfree(hctxs[i]); | |
2028 | hctxs[i] = NULL; | |
2029 | break; | |
2030 | } | |
e4043dcf | 2031 | |
0d2602ca | 2032 | atomic_set(&hctxs[i]->nr_active, 0); |
f14bbe77 | 2033 | hctxs[i]->numa_node = node; |
320ae51f | 2034 | hctxs[i]->queue_num = i; |
868f2f0b KB |
2035 | |
2036 | if (blk_mq_init_hctx(q, set, hctxs[i], i)) { | |
2037 | free_cpumask_var(hctxs[i]->cpumask); | |
2038 | kfree(hctxs[i]); | |
2039 | hctxs[i] = NULL; | |
2040 | break; | |
2041 | } | |
2042 | blk_mq_hctx_kobj_init(hctxs[i]); | |
320ae51f | 2043 | } |
868f2f0b KB |
2044 | for (j = i; j < q->nr_hw_queues; j++) { |
2045 | struct blk_mq_hw_ctx *hctx = hctxs[j]; | |
2046 | ||
2047 | if (hctx) { | |
2048 | if (hctx->tags) { | |
2049 | blk_mq_free_rq_map(set, hctx->tags, j); | |
2050 | set->tags[j] = NULL; | |
2051 | } | |
2052 | blk_mq_exit_hctx(q, set, hctx, j); | |
2053 | free_cpumask_var(hctx->cpumask); | |
2054 | kobject_put(&hctx->kobj); | |
2055 | kfree(hctx->ctxs); | |
2056 | kfree(hctx); | |
2057 | hctxs[j] = NULL; | |
2058 | ||
2059 | } | |
2060 | } | |
2061 | q->nr_hw_queues = i; | |
2062 | blk_mq_sysfs_register(q); | |
2063 | } | |
2064 | ||
2065 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, | |
2066 | struct request_queue *q) | |
2067 | { | |
66841672 ML |
2068 | /* mark the queue as mq asap */ |
2069 | q->mq_ops = set->ops; | |
2070 | ||
868f2f0b KB |
2071 | q->queue_ctx = alloc_percpu(struct blk_mq_ctx); |
2072 | if (!q->queue_ctx) | |
c7de5726 | 2073 | goto err_exit; |
868f2f0b KB |
2074 | |
2075 | q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)), | |
2076 | GFP_KERNEL, set->numa_node); | |
2077 | if (!q->queue_hw_ctx) | |
2078 | goto err_percpu; | |
2079 | ||
2080 | q->mq_map = blk_mq_make_queue_map(set); | |
2081 | if (!q->mq_map) | |
2082 | goto err_map; | |
2083 | ||
2084 | blk_mq_realloc_hw_ctxs(set, q); | |
2085 | if (!q->nr_hw_queues) | |
2086 | goto err_hctxs; | |
320ae51f | 2087 | |
287922eb | 2088 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 2089 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f JA |
2090 | |
2091 | q->nr_queues = nr_cpu_ids; | |
320ae51f | 2092 | |
94eddfbe | 2093 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 2094 | |
05f1dd53 JA |
2095 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
2096 | q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE; | |
2097 | ||
1be036e9 CH |
2098 | q->sg_reserved_size = INT_MAX; |
2099 | ||
2849450a | 2100 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
2101 | INIT_LIST_HEAD(&q->requeue_list); |
2102 | spin_lock_init(&q->requeue_lock); | |
2103 | ||
07068d5b JA |
2104 | if (q->nr_hw_queues > 1) |
2105 | blk_queue_make_request(q, blk_mq_make_request); | |
2106 | else | |
2107 | blk_queue_make_request(q, blk_sq_make_request); | |
2108 | ||
eba71768 JA |
2109 | /* |
2110 | * Do this after blk_queue_make_request() overrides it... | |
2111 | */ | |
2112 | q->nr_requests = set->queue_depth; | |
2113 | ||
24d2f903 CH |
2114 | if (set->ops->complete) |
2115 | blk_queue_softirq_done(q, set->ops->complete); | |
30a91cb4 | 2116 | |
24d2f903 | 2117 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
320ae51f | 2118 | |
5778322e | 2119 | get_online_cpus(); |
320ae51f | 2120 | mutex_lock(&all_q_mutex); |
320ae51f | 2121 | |
4593fdbe | 2122 | list_add_tail(&q->all_q_node, &all_q_list); |
0d2602ca | 2123 | blk_mq_add_queue_tag_set(set, q); |
5778322e | 2124 | blk_mq_map_swqueue(q, cpu_online_mask); |
484b4061 | 2125 | |
4593fdbe | 2126 | mutex_unlock(&all_q_mutex); |
5778322e | 2127 | put_online_cpus(); |
4593fdbe | 2128 | |
320ae51f | 2129 | return q; |
18741986 | 2130 | |
320ae51f | 2131 | err_hctxs: |
868f2f0b | 2132 | kfree(q->mq_map); |
f14bbe77 | 2133 | err_map: |
868f2f0b | 2134 | kfree(q->queue_hw_ctx); |
320ae51f | 2135 | err_percpu: |
868f2f0b | 2136 | free_percpu(q->queue_ctx); |
c7de5726 ML |
2137 | err_exit: |
2138 | q->mq_ops = NULL; | |
320ae51f JA |
2139 | return ERR_PTR(-ENOMEM); |
2140 | } | |
b62c21b7 | 2141 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f JA |
2142 | |
2143 | void blk_mq_free_queue(struct request_queue *q) | |
2144 | { | |
624dbe47 | 2145 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2146 | |
0e626368 AM |
2147 | mutex_lock(&all_q_mutex); |
2148 | list_del_init(&q->all_q_node); | |
2149 | mutex_unlock(&all_q_mutex); | |
2150 | ||
0d2602ca JA |
2151 | blk_mq_del_queue_tag_set(q); |
2152 | ||
624dbe47 ML |
2153 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
2154 | blk_mq_free_hw_queues(q, set); | |
320ae51f | 2155 | } |
320ae51f JA |
2156 | |
2157 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
5778322e AM |
2158 | static void blk_mq_queue_reinit(struct request_queue *q, |
2159 | const struct cpumask *online_mask) | |
320ae51f | 2160 | { |
4ecd4fef | 2161 | WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth)); |
320ae51f | 2162 | |
67aec14c JA |
2163 | blk_mq_sysfs_unregister(q); |
2164 | ||
5778322e | 2165 | blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues, online_mask); |
320ae51f JA |
2166 | |
2167 | /* | |
2168 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
2169 | * we should change hctx numa_node according to new topology (this | |
2170 | * involves free and re-allocate memory, worthy doing?) | |
2171 | */ | |
2172 | ||
5778322e | 2173 | blk_mq_map_swqueue(q, online_mask); |
320ae51f | 2174 | |
67aec14c | 2175 | blk_mq_sysfs_register(q); |
320ae51f JA |
2176 | } |
2177 | ||
f618ef7c PG |
2178 | static int blk_mq_queue_reinit_notify(struct notifier_block *nb, |
2179 | unsigned long action, void *hcpu) | |
320ae51f JA |
2180 | { |
2181 | struct request_queue *q; | |
5778322e AM |
2182 | int cpu = (unsigned long)hcpu; |
2183 | /* | |
2184 | * New online cpumask which is going to be set in this hotplug event. | |
2185 | * Declare this cpumasks as global as cpu-hotplug operation is invoked | |
2186 | * one-by-one and dynamically allocating this could result in a failure. | |
2187 | */ | |
2188 | static struct cpumask online_new; | |
320ae51f JA |
2189 | |
2190 | /* | |
5778322e AM |
2191 | * Before hotadded cpu starts handling requests, new mappings must |
2192 | * be established. Otherwise, these requests in hw queue might | |
2193 | * never be dispatched. | |
2194 | * | |
2195 | * For example, there is a single hw queue (hctx) and two CPU queues | |
2196 | * (ctx0 for CPU0, and ctx1 for CPU1). | |
2197 | * | |
2198 | * Now CPU1 is just onlined and a request is inserted into | |
2199 | * ctx1->rq_list and set bit0 in pending bitmap as ctx1->index_hw is | |
2200 | * still zero. | |
2201 | * | |
2202 | * And then while running hw queue, flush_busy_ctxs() finds bit0 is | |
2203 | * set in pending bitmap and tries to retrieve requests in | |
2204 | * hctx->ctxs[0]->rq_list. But htx->ctxs[0] is a pointer to ctx0, | |
2205 | * so the request in ctx1->rq_list is ignored. | |
320ae51f | 2206 | */ |
5778322e AM |
2207 | switch (action & ~CPU_TASKS_FROZEN) { |
2208 | case CPU_DEAD: | |
2209 | case CPU_UP_CANCELED: | |
2210 | cpumask_copy(&online_new, cpu_online_mask); | |
2211 | break; | |
2212 | case CPU_UP_PREPARE: | |
2213 | cpumask_copy(&online_new, cpu_online_mask); | |
2214 | cpumask_set_cpu(cpu, &online_new); | |
2215 | break; | |
2216 | default: | |
320ae51f | 2217 | return NOTIFY_OK; |
5778322e | 2218 | } |
320ae51f JA |
2219 | |
2220 | mutex_lock(&all_q_mutex); | |
f3af020b TH |
2221 | |
2222 | /* | |
2223 | * We need to freeze and reinit all existing queues. Freezing | |
2224 | * involves synchronous wait for an RCU grace period and doing it | |
2225 | * one by one may take a long time. Start freezing all queues in | |
2226 | * one swoop and then wait for the completions so that freezing can | |
2227 | * take place in parallel. | |
2228 | */ | |
2229 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2230 | blk_mq_freeze_queue_start(q); | |
f054b56c | 2231 | list_for_each_entry(q, &all_q_list, all_q_node) { |
f3af020b TH |
2232 | blk_mq_freeze_queue_wait(q); |
2233 | ||
f054b56c ML |
2234 | /* |
2235 | * timeout handler can't touch hw queue during the | |
2236 | * reinitialization | |
2237 | */ | |
2238 | del_timer_sync(&q->timeout); | |
2239 | } | |
2240 | ||
320ae51f | 2241 | list_for_each_entry(q, &all_q_list, all_q_node) |
5778322e | 2242 | blk_mq_queue_reinit(q, &online_new); |
f3af020b TH |
2243 | |
2244 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2245 | blk_mq_unfreeze_queue(q); | |
2246 | ||
320ae51f JA |
2247 | mutex_unlock(&all_q_mutex); |
2248 | return NOTIFY_OK; | |
2249 | } | |
2250 | ||
a5164405 JA |
2251 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
2252 | { | |
2253 | int i; | |
2254 | ||
2255 | for (i = 0; i < set->nr_hw_queues; i++) { | |
2256 | set->tags[i] = blk_mq_init_rq_map(set, i); | |
2257 | if (!set->tags[i]) | |
2258 | goto out_unwind; | |
2259 | } | |
2260 | ||
2261 | return 0; | |
2262 | ||
2263 | out_unwind: | |
2264 | while (--i >= 0) | |
2265 | blk_mq_free_rq_map(set, set->tags[i], i); | |
2266 | ||
a5164405 JA |
2267 | return -ENOMEM; |
2268 | } | |
2269 | ||
2270 | /* | |
2271 | * Allocate the request maps associated with this tag_set. Note that this | |
2272 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
2273 | * will be updated to reflect the allocated depth. | |
2274 | */ | |
2275 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2276 | { | |
2277 | unsigned int depth; | |
2278 | int err; | |
2279 | ||
2280 | depth = set->queue_depth; | |
2281 | do { | |
2282 | err = __blk_mq_alloc_rq_maps(set); | |
2283 | if (!err) | |
2284 | break; | |
2285 | ||
2286 | set->queue_depth >>= 1; | |
2287 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
2288 | err = -ENOMEM; | |
2289 | break; | |
2290 | } | |
2291 | } while (set->queue_depth); | |
2292 | ||
2293 | if (!set->queue_depth || err) { | |
2294 | pr_err("blk-mq: failed to allocate request map\n"); | |
2295 | return -ENOMEM; | |
2296 | } | |
2297 | ||
2298 | if (depth != set->queue_depth) | |
2299 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
2300 | depth, set->queue_depth); | |
2301 | ||
2302 | return 0; | |
2303 | } | |
2304 | ||
f26cdc85 KB |
2305 | struct cpumask *blk_mq_tags_cpumask(struct blk_mq_tags *tags) |
2306 | { | |
2307 | return tags->cpumask; | |
2308 | } | |
2309 | EXPORT_SYMBOL_GPL(blk_mq_tags_cpumask); | |
2310 | ||
a4391c64 JA |
2311 | /* |
2312 | * Alloc a tag set to be associated with one or more request queues. | |
2313 | * May fail with EINVAL for various error conditions. May adjust the | |
2314 | * requested depth down, if if it too large. In that case, the set | |
2315 | * value will be stored in set->queue_depth. | |
2316 | */ | |
24d2f903 CH |
2317 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
2318 | { | |
205fb5f5 BVA |
2319 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
2320 | ||
24d2f903 CH |
2321 | if (!set->nr_hw_queues) |
2322 | return -EINVAL; | |
a4391c64 | 2323 | if (!set->queue_depth) |
24d2f903 CH |
2324 | return -EINVAL; |
2325 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
2326 | return -EINVAL; | |
2327 | ||
f9018ac9 | 2328 | if (!set->ops->queue_rq || !set->ops->map_queue) |
24d2f903 CH |
2329 | return -EINVAL; |
2330 | ||
a4391c64 JA |
2331 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
2332 | pr_info("blk-mq: reduced tag depth to %u\n", | |
2333 | BLK_MQ_MAX_DEPTH); | |
2334 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
2335 | } | |
24d2f903 | 2336 | |
6637fadf SL |
2337 | /* |
2338 | * If a crashdump is active, then we are potentially in a very | |
2339 | * memory constrained environment. Limit us to 1 queue and | |
2340 | * 64 tags to prevent using too much memory. | |
2341 | */ | |
2342 | if (is_kdump_kernel()) { | |
2343 | set->nr_hw_queues = 1; | |
2344 | set->queue_depth = min(64U, set->queue_depth); | |
2345 | } | |
868f2f0b KB |
2346 | /* |
2347 | * There is no use for more h/w queues than cpus. | |
2348 | */ | |
2349 | if (set->nr_hw_queues > nr_cpu_ids) | |
2350 | set->nr_hw_queues = nr_cpu_ids; | |
6637fadf | 2351 | |
868f2f0b | 2352 | set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *), |
24d2f903 CH |
2353 | GFP_KERNEL, set->numa_node); |
2354 | if (!set->tags) | |
a5164405 | 2355 | return -ENOMEM; |
24d2f903 | 2356 | |
a5164405 JA |
2357 | if (blk_mq_alloc_rq_maps(set)) |
2358 | goto enomem; | |
24d2f903 | 2359 | |
0d2602ca JA |
2360 | mutex_init(&set->tag_list_lock); |
2361 | INIT_LIST_HEAD(&set->tag_list); | |
2362 | ||
24d2f903 | 2363 | return 0; |
a5164405 | 2364 | enomem: |
5676e7b6 RE |
2365 | kfree(set->tags); |
2366 | set->tags = NULL; | |
24d2f903 CH |
2367 | return -ENOMEM; |
2368 | } | |
2369 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
2370 | ||
2371 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
2372 | { | |
2373 | int i; | |
2374 | ||
868f2f0b | 2375 | for (i = 0; i < nr_cpu_ids; i++) { |
f42d79ab | 2376 | if (set->tags[i]) |
484b4061 JA |
2377 | blk_mq_free_rq_map(set, set->tags[i], i); |
2378 | } | |
2379 | ||
981bd189 | 2380 | kfree(set->tags); |
5676e7b6 | 2381 | set->tags = NULL; |
24d2f903 CH |
2382 | } |
2383 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
2384 | ||
e3a2b3f9 JA |
2385 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
2386 | { | |
2387 | struct blk_mq_tag_set *set = q->tag_set; | |
2388 | struct blk_mq_hw_ctx *hctx; | |
2389 | int i, ret; | |
2390 | ||
2391 | if (!set || nr > set->queue_depth) | |
2392 | return -EINVAL; | |
2393 | ||
2394 | ret = 0; | |
2395 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
2396 | if (!hctx->tags) |
2397 | continue; | |
e3a2b3f9 JA |
2398 | ret = blk_mq_tag_update_depth(hctx->tags, nr); |
2399 | if (ret) | |
2400 | break; | |
2401 | } | |
2402 | ||
2403 | if (!ret) | |
2404 | q->nr_requests = nr; | |
2405 | ||
2406 | return ret; | |
2407 | } | |
2408 | ||
868f2f0b KB |
2409 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) |
2410 | { | |
2411 | struct request_queue *q; | |
2412 | ||
2413 | if (nr_hw_queues > nr_cpu_ids) | |
2414 | nr_hw_queues = nr_cpu_ids; | |
2415 | if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues) | |
2416 | return; | |
2417 | ||
2418 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2419 | blk_mq_freeze_queue(q); | |
2420 | ||
2421 | set->nr_hw_queues = nr_hw_queues; | |
2422 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
2423 | blk_mq_realloc_hw_ctxs(set, q); | |
2424 | ||
2425 | if (q->nr_hw_queues > 1) | |
2426 | blk_queue_make_request(q, blk_mq_make_request); | |
2427 | else | |
2428 | blk_queue_make_request(q, blk_sq_make_request); | |
2429 | ||
2430 | blk_mq_queue_reinit(q, cpu_online_mask); | |
2431 | } | |
2432 | ||
2433 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2434 | blk_mq_unfreeze_queue(q); | |
2435 | } | |
2436 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); | |
2437 | ||
676141e4 JA |
2438 | void blk_mq_disable_hotplug(void) |
2439 | { | |
2440 | mutex_lock(&all_q_mutex); | |
2441 | } | |
2442 | ||
2443 | void blk_mq_enable_hotplug(void) | |
2444 | { | |
2445 | mutex_unlock(&all_q_mutex); | |
2446 | } | |
2447 | ||
320ae51f JA |
2448 | static int __init blk_mq_init(void) |
2449 | { | |
320ae51f JA |
2450 | blk_mq_cpu_init(); |
2451 | ||
add703fd | 2452 | hotcpu_notifier(blk_mq_queue_reinit_notify, 0); |
320ae51f JA |
2453 | |
2454 | return 0; | |
2455 | } | |
2456 | subsys_initcall(blk_mq_init); |